arXiv:1604.00607v1 [cs.IT] 3 Apr 2016 o 03B360 n h tt ao cec n Technolog and Science u Major 2016ZX03001020-006). Program) State No. (973 (Grant the China Projects of and Program 2013CB336600, 2014 Research No. No. Tech Basic Grant High under National National China the of the Program 61361166005, Development No. and search Grant under China of eie e.0,21;Acpe a.1,2016. 19, Mar. Accepted 2016; 04, Feb. Revised mail: hngn ag(-al gagie.r)i ihteInter Ch R. the P. with Shaanxi, is USA. (e-mail 710119, PA, [email protected]) Prussia, of Xi’an Precisi Tr Li (e-mail: King and Communications, Sciences, of Wang Optics of Xuelong Laboratory of Chonggang Institute Key Academy Xi’an China. State Chinese Photonics, (OPTIMAL), and Telecommunications, Learning Optics and Beijin and Analysis Education, of Posts [email protected] Ministry for of Communications Wireless [email protected] otiuost h xlsv oietafi rwhi h ra the is It growth primary increas traffic [1]. the mobile 30 will of 2000 explosive One traffic the nearly [2]. in to data 2019 contributors was and Internet mobile 2014 global global between and tenfold the entire nearly percent that mobile the 69 predicted the of is grew that size shows 2014 the Cisco times in from traffic report A data years. few past introduc th are and test discussed, trial are and well. development C-RANs testbed for of security soci progress mining, layer data physical big defined and cache, software radio, edge cognitive investig device-to-device, of aware future involvement spur the lay area to which presented research upper in are the the challenges of and in extensiveness issues optimization the given and a Additionally, layer; allocation physical resource the in radio estimation coll channel large-scale and compression, processing, fronthaul C-RA the in as: techniques classified key state-of-the-art summa The discussed. comprehensively and splits are functional characteristics i different corresponding open with and architectures system techniques, C-RAN This The key of community. architectures, advances research system recent and including the surveys industry theoreti comprehensively the efficiency be paper significant both have spectral its C-RANs by advantages, high advocated network by potential and provide Motivated access gains to efficiency. performance radio shown energy cloud been and the has expenditures, (C-RAN) operating and in eoreallocation. resource tion, chan processing, collaborative large-scale compression, EECMUIAIN UVY UOIL,VL ,N.Y MON. Y, NO. X, VOL. TUTORIALS, & SURVEYS COMMUNICATIONS IEEE eetAvne nCodRdoAcs Networks: Access Radio Cloud in Advances Recent ue eg(e-mail: Peng Mugen hswr a upre npr yteNtoa aua Scien Natural National the by part in supported was work This aucitrcie e.2,21;FrtRvsdDc 3 20 23, Dec. Revised First 2015; 24, Sep. received Manuscript oiedt rfchsbe rwn xoetal vrthe over exponentially growing been has traffic data Mobile Terms Index Abstract ytmAcietrs e ehius n Open and Techniques, Key Architectures, System [email protected] A rmsn aaimt euebt capital both reduce to paradigm promising a —As Codrdoacs ewr CRN,fronthaul (C-RAN), network access radio —Cloud r ihteKyLbrtr fUniversal of Laboratory Key the with are ) .I I. ue eg ahaSn uln i hnogMo n Chongg and Mao, Zhendong Li, Xuelong Sun, Yaohua Peng, Mugen swt h etrfrOTclIMagery OPTical for Center the with is ) [email protected] NTRODUCTION ,adZedn a (e-mail: Mao Zhendong and ), ,You u (e- Sun Yaohua ), nMechanics, on eFoundation ce University g e estima- nel AA01A701, 5 Second 15; ooyRe- nology aborative drGrant nder network, Special y n the and ations, dthe nd sare Ns open , Issues ansient Digital ssues. das ed rized pid 2016 ina. cal al- er. en s, e e : optblt.I atclr -Ascnrdc h otof backward cost and the forward reduce SE, including can and C-RANs, C-RANs EE particular, of In high requirements compatibility. cost, and reduced vision the the on ing [10]. plane data in-ph and specificat plane digitized includes control bidirectional and for transmission, and (I/Q) sup- rate quadrature which via bit and protocol, constant RRHs (CPRI) a with interface ports communicates radio public pool common BBU intelli The and (LSCP (CRRA), networking. processing allocation resource BB collaborative radio virtualized cooperative large-scale the while provides spots, hot pool in functi frequency capacity pool radio high BBU with support RRHs a location. as centralized clustered a hea (BBUs) in radio units remote baseband traditiona distributed the and parts: 1, (RRHs), two Fig. into in decoupled virtua shown is as via BS C-RAN resources general computing a In and ization. storage pri sharing cloud centralized of the ciple on improv based to (EE) efficiency potential energy the informa and has the that and industries network technology wireless tion the both from technologies [7]–[9]. systems communication ra wireless bit advanced equipment the transmission expenditu in energy-efficiency the operating high and and providing capital operators while both the curtails that both evolvedvendors by an as paradigm recognized been system acc recently radio has e cloud (C-RAN) and the network Accordingly, resource [6]. optimize [5], to consumption as i ergy BSs so low of server functions and cloud base the centralized move SE by the to consumed appealing high is is it of energy (BSs), stations requirements most Since future consumption. the power meet and netw reduce cost to architectures network wireless intelligent telecommunica- [4]. the devices from mobile and was 200 infrastructure 25% In tion over about was [3]. which industries of profit information gigawatt, communication by operators’ and consumption (IT) mobile power technology average of annual increase the 2012, slow especiall expenditure the problematic, operating becomes with of has growth (SE) consumption rapid efficiency energy the and spectral hand, high other offer the On to acce expected wireless are t next-generation combat networks the To traffic demand, future. foreseeable growing mobile the rapidly in in continue increases to applications projected mobile sharp bandwidth-hungry these from mob (particularly and on running devices, applications smart social mobile of proliferation h ieesidsr n eerhcmuiyaework- are community research and industry wireless The emerging of combination breakthrough a is C-RAN The advanced for calling increasingly is world wireless The n Wang ang orking are ) SE e gent ions res, ons nto ase ess tes ile he ds n- n- ss U ), l- y 1 - l . IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 2

complexity, and imperfections in the channel state information

Virtualized BBU Pool (CSIs) collected at the BBU pool from all connected RRHs. Core network %DFNKDXO Particularly, in the uplink, each RRH compresses its received signal for transmission to the central BBU pool via the con- :LUHOHVV &DSDFLW\ IURQWKDXO OLPLWHG:LUHG strained fronthaul link. The central BBU pool then performs IURQWKDXO   joint decoding of the data streams of all user equipments  :LUHOHVV

 EDFNKDXO (UEs), based on the received compressed signals. Compression  8( 55+ over the constrained fronthaul and large-scale decoding pose %%8SRRO  significant technical challenges to be tackled for a successful  1HWZRUN rollout and commercial operations. In the downlink, each UE 0HGLXP$FFHVV :LUHG &RQWURO 55+ EDFNKDXO receives signals from adjacent RRHs and interfering RRHs, %DVHEDQG 5DGLR)UHTXHQF\ 3URFHVVLQJ which include interference and useful signal. Precoding at &35,LQWHUIDFH &35,LQWHUIDFH the BBU pool, along with channel estimation and detection )URQWKDXO Fig. 1. A general C-RAN system at the UE, pose the technical challenges. Additionally, radio resource management is a key functionality run at the BBU pool to decrease the interference and achieve the potential performance gains. the deployment owing to the possibility to substitute full- Recently, diverse problems and corresponding solutions for fledged dense BSs with RRHs. The RRH deployment results in C-RANs have been intensively elaborated from the aspects of reducing the construction space and saving energy consump- system architecture, fronthaul compression, LSCP, and CRRA, tions of dense BSs. In [11], a quantitative analysis of the cost allowing feasible designs and operations of energy efficient was presented in C-RANs, and it was shown that C-RANs and spectrum efficient C-RANs. To systematically show the can lead to a capital expenditure reduction with 10% to 15% advances in C-RANs, several survey works have been done per kilometer comparing with traditional long term evolution [13]–[18]. The progress on the centralization and virtualization (LTE) networks. Meanwhile, C-RANs enable the flexible for C-RANs is introduced in [13], where the performances allocation of scarce radio and computing resources across of several key technologies like uplink coordinated multiple all RRHs centrally processed by the same BBU pool, which point (CoMP) are verified, and a general purpose platform reaps statistical multiplexing gains due to the load balance (GPP) based C-RAN test bed is built, showing almost the of adjacent RRHs for the un-uniformly traffic distribution. same performance as traditional systems. While in [14], a In [12], the equal per-cell rate was analyzed for half-duplex high-level overview of C-RANs is provided, including the and full-duplex operations in C-RANs, and the simulation architecture, transport network techniques, virtualization and results demonstrated that significant performance gains can challenges. A novel logical structure of C-RANs comprising be achieved by both operation modes compared with the cor- physical plane, control plane, and service plane is presented in responding single-cell processing approaches under unlimited [15], along with a coordinated user scheduling algorithm and fronthaul capacity. Furthermore, C-RANs simplify the cellular parallel optimum precoding scheme. In [16], a comprehensive network upgrading and maintenance due to the centralized survey related to recent advances in fronthaul-constrained C- architecture. At the BBU pool, baseband processing can be RANs is discussed, mainly focusing on potential solutions to implemented using software radio technology based on open enhance SE and EE. In addition, the authors in [17] reviewed IT architectures, making the systems upgrading to different the key fronthaul compression techniques for the uplink and standards possible without hardware upgrades. With C-RANs, downlink of a C-RAN, together with simulation results to mobile operators can quickly deploy RRHs to expand and validate the performance improvement desired from the imple- make upgrades to their cellular network. Mobile operators only mentation of multiterminal fronthaul compression. In [18], the need to install new RRHs and connect them to the BBU pool developments of resource allocation in heterogeneous cloud in order to expand the network coverage or split the cell to radio access networks (H-CRANs) are investigated, and three improve SE. resource allocation schemes are proposed, i.e., coordinated Motivated by the potential significant benefits, C-RANs scheduling, hybrid backhauling, and multicloud association. have been advocated by both mobile operators (e.g., France Although initial efforts have been made on the survey of Telecom/Orange, NTT DoCoMo, Telefonica, and China Mo- C-RANs, limitations exist in the previous works that the bile) and equipment vendors (e.g., Cisco, Intel, IBM, Huawei, content is only related to the practical implementation and ZTE, Ericsson, and so on). Despite the mentioned attractive deployment aspects in [13], or only a small part of current features, C-RANs also come with their own technological research achievements on C-RANs are specified in [15], or challenges. For example, deploying a large number of RRHs only a specific topic, which is relative to H-CRANs but not entails significant interference when LSCP works inefficiently the general C-RANs, is concentrated on in [18]. Considering due to the constrained fronthaul, which presents a bottleneck these shortcomings and the ongoing research activities, a more to the capacity of C-RANs. In practice, in order to mitigate comprehensive survey framework for incorporating the basics interference, it is necessary to exploit the large-scale pre- and the latest achievements on the system architecture and coding and decoding schemes at the BBU pool that takes key techniques across different layers over the air interface into account the fronthaul capacity limitations, implementation seems timely and significant. As shown in Table I, compared IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 3

TABLE I with the existing survey articles, a more comprehensive survey SUMMARYOF EXISTING SURVEY ARTICLES framework for recent advances in system architectures and key techniques to improve SE and EE is proposed in this paper. Aspects Survey Contributions Papers Specifically, this envisioned paper surveys the state-of-the-art [13] Initial results from the field trials of the central- ization and achievements on the prototype devel- system architectures, key technologies on the advanced LSCP, Comprehensive opment of virtualized C-RANs. Surveys fronthaul compression, and channel estimation in PHY, as well [14] A high level overview of traditional C-RANs in- as CRRA in the upper layers. Additionally, given the exten- cluding the architecture, the transport network, the virtualization, and so on. siveness of the research area, open issues and challenges are [15] A logical structure of C-RANs comprising physical presented to spur future investigations. The aims of this survey plane, control plane, and service plane. [16] Recent advances in fronthaul-constrained C-RANs, are to fill the research gaps found in the previous publications focusing on potential solutions to enhancing SE [13]–[18] as stated above. In this way, the contributions of this and EE. This A comprehensive survey of state-of-the-art system paper are fourfold as summarized below: Article architectures, key technologies on the advanced LSCP in PHY, and CRRA in the upper layers. 1) A comprehensive survey of system architectures is pre- Fronthaul [17] A review of quantization based fronthaul compres- Compression and sion techniques for both the uplink and downlink sented, which is divided into the architectures proposed Quantization of a C-RAN. by the industry and the architectures proposed by the This A comprehensive survey of latest fronthaul com- Article pression techniques, including quantization based academia. Particularly, architectures proposed by the fronthaul compression, compressive sensing based industry are illustrated based on different functional compression, and spatial filtering. Large-Scale This A comprehensive survey of precoding techniques splits, in which the tradeoff between implementation Collaborative Article with full CSIs, precoding techniques with partial complexity and performance gains is concerned. The Processing CSIs, and sparse precoding. Channel Estima- This An overview of channel estimation techniques re- system architecture evolution to H-CRANs and fog tion and Training Article cently proposed for C-RANs, giving useful guide- computing based radio access networks (F-RANs) is Design lines on the new requirements when employing the conventional channel estimation approaches in C- highlighted in the research community. RANs. 2) A comprehensive survey of key techniques in PHY is Radio Resource [18] An investigation of developments in centralized ap- presented, including the fronthaul compression, LSCP, Allocation and proaches to resource allocation only in the special Optimization H-CRANs but not for a general C-RAN. and channel estimation. The principle, challenges and This A full survey of optimization approaches to static the technical solutions of these key techniques in PHY Article CRRA without considering QSI and dynamic CRRA with queue-awareness in C-RANs, covering are summarized, and the contributions and researching both classic approaches and game model based results of relative literatures are systematically elabo- approaches. rated. Particularly, the quantization, compressive sensing (CS), and spatial filtering for the fronthaul compres- sion in both uplink and downlink are summarized. Section V summarizes recent advances of channel estimation Meanwhile, both the linear LSCP with/without perfect and training design solutions. The static and dynamic CRRA CSIs and the nonlinear sparse LSCP are presented. In are surveyed in VI. The current challenges and open issues addition, the superimposed training, segment training, are shown in Section VII, prior to the conclusion in Section and semi-blind channel estimation approaches for C- VIII. For convenience, all abbreviations are listed in Table II. RANs are surveyed. 3) The CRRA for C-RANs are comprehensively summa- II. SYSTEM ARCHITECTURES rized, including the static CRRA without considering In recent years, C-RANs have attracted much attention queue state information (QSI) and the dynamic CRRA of both industries and the academia due to their significant with queue-awareness. In particular, the optimization benefits to meet the enormous demand for data traffic and approaches for the static CRRA are classified into classic meanwhile decrease the capital and operational expenditures. non-convex optimization approaches and game model In the industry, many operators and equipment vendors have based approaches. The dynamic CRRA is surveyed into actively expressed their own opinions on C-RAN architectures the equivalent rate approach, the Lyapunov optimization in the form of technical reports or literatures. While in the approach, and the Markov decision process approach. academia, different solutions have been proposed for the 4) The future challenges and open issues related to C- enhancements of C-RANs. In this section, the general system RANs are identified, as edge cache, big data mining, architecture of C-RANs is firstly introduced, and then the social-aware device-to-device (D2D) communication, specific architectures in the industry and academia are briefly cognitive radio (CR), software defined network (SDN), surveyed. physical layer security, and trial test are vital for the further development of C-RANs. A. C-RAN General Architectures The rest of this paper is organized as follows: Section The evolution of cellular BSs in the mobile communication II surveys the system architectures of C-RANs from the system is shown in Fig. ??. In 1G and 2G cellular networks, literatures related to the industry and research community. The radio and baseband processing functionalities are integrated uplink and downlink compression to decrease the capacity inside a BS. While in 3G and 4G cellular networks, a BS is requirements of fronthaul are summarized in Section III. In divided into a RRH and a BBU. In this BS architecture, the Section IV, the LSCP techniques for C-RANs are surveyed. location of the BBU can be far from the RRH for a low site IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 4

rental and the convenience of maintenance. In the 4G beyond, BBUs are migrated into a BBU pool, which is virtualized and TABLE II shared by different cell sites, and RRHs are connected to the SUMMARY OF ABBREVIATIONS BBU pool via the fronthaul links, which results in a general 3G 3rd generation C-RAN architecture. 4th 4G generation By the centralized LSCP in C-RANs, several benefits can 5G 5th generation ACL radio access link be achieved. First, a huge energy consumption incurred by air BBU baseband unit conditioning can be saved owing to the reduction of BS sites. BS base station Second, the operation and maintenance cost can be decreased CA carrier aggregation CoMP coordinated multiple point because BBUs are placed in several big rooms. Third, the CPRI common public radio interface LSCP technique can be implemented more easily due to the CR cognitive radio centralized structure than the traditional cellular system, which C-RAN cloud radio access network CRRA cooperative radio resource allocation significantly improves SE. Fourth, C-RANs can be endowed CS compressive sensing with the capability of load balancing in the BBU pool, and CSI channel state information thus adapts to the non-uniform traffic. CWDM coarse wavelength division multiplexing D2D device-to-device D-RAN distributed radio access network B. C-RAN Architectures Proposed by Industries DAS distributed antenna system DPR decompress-process-and-recompress As shown in Fig. 2, the history of C-RANs in the industry DWDM dense wavelength division multiplexing can be traced back to 2010, in which the concept of C-RANs ECF estimate-compress-forward EE energy efficiency is firstly proposed with the name of wireless network cloud EPON Ethernet PON (WNC) by IBM to decrease networking cost and obtain more F-RAN fog computing based radio access network flexible network capabilities [2]. Then this concept is further GSBF group sparse beamforming GSM global system for mobile communications exploited by China Mobile Research Institute in 2011 [3], in H-CRAN heterogeneous cloud radio access network which the C-RAN architecture is elaborated along with the I/Q in-phase and quadrature technology trends and feasibility analysis. To deal with the LMMSE linear minimum-mean-square-error LS least-squared fiber scarcity, ZTE proposed different solutions, including en- LSCP large-scale collaborative processing hanced fiber connection, colored fiber connection, and optical LTE long term evolution transport network bearer, which can use a very few fibers to MA multiple access MAC medium access control meet the bearer requirements of C-RANs [4]. An efficient and MARN multiple access relay networks scalable GPP based BBU pool architecture is proposed by MBS macro BS Intel, and it can utilize computation resource as needed [5]. MDP Markov decision process MF multiplex-and-forward Leveraging virtualization techniques, the concept of cloud BSs NG-PON next-generation passive optical network is proposed by Alcatel-Lucent, which requires less processing NGMN next generation mobile networks resources without degrading system performances [6]. In 2013, MIMO multiple input multiple output ML maximum likelihood NTT DoCoMo began developing the advanced centralized C- MU-MIMO multiuser MIMO RAN architecture for its future LTE-Advanced mobile system PE polynomial expansion [8]. Also in 2013, the RAN-as-a-Service (RANaaS) concept PHY physical QoE quality of experience is emphasized by Telecom Italia, which provides the benefits QoS quality of service of the C-RAN architecture with more flexibility [7]. While in QSI queue state information 2014, the white paper of Liquid Radio is released by Nokia RANaaS radio access network as a service RF radio frequency Networks, in which the centralized C-RAN is taken as an RRH remote radio head efficient way to enhance the network utilization [9]. RRM radio resource management SDF software defined fronthaul China Mobile: C-RAN SDN software defined network NTT DoCoMo: white paper [3] Intel: SE spectrum efficiency GPP based BBU pool Advanced C-RAN architecture [8] SINR signal-to-interference-plus-noise ratio architecture [5] Telecom Italia: SQNR signal-to-quantization-noise-ratio IBM: WNC [2]2] RANaaS [7] SS superimposed-segment TD-LTE time division-long term evolution      TD-SCDMA time division-synchronization code division multiple access TDMA time division multiple access Nokia: ZTE: Alcatel-lucent: Liquid radio white paper [9] TWDM time wavelength division multiplexing Enhanced fiber Cloud base stations [6] UE user equipment connection [4] WDM wavelength-division multiplexing Fig. 2. The milestone of C-RAN architectures in the industry. WSR weighted sum rate WMMSE weighted minimum mean square error WNC wireless network cloud 1) Functional Split of C-RANs: To take a tradeoff between implementing complexity and achieved LSCP gains, one of the key differences among those architectures advocated by the industry lies in the degree of functional split between IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 5

RRHs and the BBU pool, as illustrated in Fig. 3. For which has been defined with several options, such as op- example, China Mobile Research Institute proposed fully- tion 1 (0.6144 Gbps), 2 (1.2288 Gbps), 3 (2.4576 Gbps), centralized and partially-centralized architectures [3]. The 6 (6.144 Gbps) and 7 (9.83 Gbps) with different carrying fully-centralized architecture, like Solution 1 in Fig. 3, shows a technologies [19]. Note that CPRI is not a standard, but an typical approach to deploy the C-RAN , in which the functions industry agreement. In [3], four solutions for the fronthaul in PHY, medium access control (MAC) and network layers are considered: dark fiber, wavelength-division multiplexing are all moved to the BBU pool. The main benefit of the fully- (WDM) based PON, unified fixed radio access, and mobile centralized architecture is that almost no digital processing radio access. Dark fiber is suitable when there are a lot of devices are required at RRHs, potentially making them very spare fiber resources, while the WDM based PON is preferred small and cheap, while the required data rate of fronthaul when fiber resources are limited, especially in the access ring. links for I/Q forwarding is comparatively high. The fully- In [20], a novel design algorithm of the access network with centralized architecture can achieve the maximal LSCP gains, a ring topology under the C-RAN architecture is proposed. however, this kind of gains is achieved at the expense of Given the requirements of the fronthaul network segment, a maximal heavy burden on fronthaul links because all I/Q in [21], a self-seeded reflective semiconductor optical amplifier signals of all UEs must be sent to the BBU pool timely. based technology for the WDM based PON is regarded as In [13], the authors report that a required fronthaul capacity an attractive solution to C-RANs. An unified architecture for under the fully-centralized architecture is 10 Gbps for the LTE the fronthaul link, employing ring-based PON, is proposed system with eight receive antennas with 20 MHz frequency in [22]. Considering the scarce fiber resources in some areas, bandwidth. ZTE Company introduced enhanced fiber connection and While in the partially-centralized architecture, like Solution colored fiber connection as the bearer network solutions of 2 in Fig. 3, functions in PHY are integrated into RRHs, C-RANs, in which only a few fibers are required to construct with other functions in upper layers incorporated at the BBU an independent C-RAN [4]. Specifically, the enhanced fiber pool. Although the heavy burden on fronthaul links is allevi- connection provide a low bit rate transmission up to 10 Mbps, ated in the partially-centralized architecture, LSCP can not while the colored fiber connection can provide a high bit rate be efficiently supported because functions in PHY remain transmission up to 1 Gbps. in the individual RRH, and only the traditional distributed For the WDM based PON, there are many kinds of COMP gain can be achieved. Obviously, the required fronthaul WDM techniques. The passive coarse wavelength division capacity and signal processing capability in the BBU pool multiplexing (CWDM) is a cost efficient optical multiplexing decrease significantly when the functional split is shifted to technology, however, the fixed wavelength assignment and the MAC layer. limited channel numbers in commercial CWDM solutions Furthermore, in [8], NTT DoCoMo Company has intro- render it no much attraction. To tackle this problem, the dense duced an advanced C-RAN architecture to utilize carrier aggre- wavelength division multiplexing (DWDM) is an attractive gation (CA) between marco and small cell carriers, in which candidate, which offers dynamic management of the channel most functions in PHY and MAC layers are implemented in assignment. Accurately, fronthaul links need to meet strict RRHs, and only upper-layer functions are moved to the BBU latency and jitter requirements in order to synchronize the pool. The Solution 3 in Fig. 3 shows that only the function of transmissions across massive RRHs. With the new NG-PON2 control plane remains in the BBU pool, which suggests only standard [4], DWDM is expected to be compliant with cost the CRRM gain can be obtained though the burden on fonthaul figures and operational needs for the capacity up to multiple can be neglected [17]. 100 Gbps. NG-PON2 supports two kinds of DWDM solutions: In contrast to the fixed functional split, the radio access 1) multiple unshared point-to-point connections via DWDM, network as a service (RANaaS) proposed by Telecom Italia and 2) multiple time division multiplexing /time division Company allows a flexible split, like Solution 4 in Fig. multiple access (TDMA) point-to-multipoint connections on 3, making it possible to choose an optimal operating point a separate set of DWDM channels (e.g., time wavelength between the full centralization and local execution [7], which division multiplexing (TWDM)) [23]. It can be anticipated that is traded off with lower LSCP gains in terms of SE and the TDMA based DWDM is a promising and preferred fronthaul requirements of the high capacity of fronthaul. The advantage carrier technology when the fronthaul capacity is unlimited of a flexible split is to reap the benefits of both extremes: and the cost of fronthaul is not a key factor for mobile significant SE/EE gains for high requirements of fronthaul, operators. and low SE/EE for capacity-free fronthaul. 2) Fronthaul Carrying Technologies: In addition, to fully boost the performance of C-RANs, the fronthaul carrying C. C-RAN Architectures Proposed by the Academia technologies are extremely important, which can be passive The general C-RAN refers to the virtualization of base optical network (PON), coax, cable, fiber, microwave, wireless station functionalities by means of cloud computing, which communication, and even millimeter wave. When considering can save the energy consumption, decrease the operation many RRHs connected to a common BBU pool in a general and maintenance cost, achieve the spatial processing gain, fully-centralized C-RAN architecture, the overall transport and adapt to the non-uniform traffic. However, the capacity- capacity for fronthaul will quickly extend into multiple 100 constrained fronthaul and the signal processing latency worsen Gbps. The typical example of fronthaul protocol is CPRI, the potential advantages of C-RANs. Meanwhile, there are IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 6

Fig. 3. Functional split of C-RANs. some emerging problems to be solved from a system ar- [26] introduced a software defined architecture for C-RANs, chitecture viewpoint for making C-RANs rollout, including which can be implemented on general purpose processors. the convergence and interworking with the existing cellular With this architecture, the C-RAN can be flexibly configured networks, the advanced massive MIMO and CR techniques to operate as different networks. efficiently working in C-RANs, and the flexibility design for 2) C-RANs with Massive MIMO and CR: To further im- meeting the technical requirements of the fifth generation prove SE and EE with advanced techniques, some researchers (5G) wireless systems. To tackle these challenges, much have proposed to use the massive MIMO and CR techniques research has been carried out to enhance the general C-RAN in C-RANs. In [27], the massive MIMO technique is involved. architecture in the research community. As shown in Fig. 4, By exploiting the proper transmit precoding or receive com- these academical works can be classified as the aspects of bining, hundreds of terminals can be simultaneously served flexible network configurations [24]–[26], the involvement of by one RRH, substantially improving the network capacity. promising techniques like massive MIMO [27] and CR [28], However, under a fully centralized C-RAN, when the number the convergence of different RANs [29] [30], as well as the of antennas at RRHs becomes large, the fronthaul has to face utilization of edge computing and storing capabilities [31]. a huge volume of data. To alleviate the fronthaul constraint while maintaining the collaborative processing gain, the func- C-RAN Architecture tionalities in PHY can be divided into a centralized part and Enhancements in Academia a distributed part. Particularly, in the downlink, precoding vectors and data symbols for the scheduled users are handled by the BBU pool, while RRHs precode the user symbols and transmit radio signals. The authors in [28] studied the Utilization of edge Flexible network Involvement of Convergence of computing and application of CR to the C-RAN, where RRHs are capable of configurations [22- massive MIMO different RANs [27] storing capabilities 24] [25] or CR [26] [28] sensing the spectrum and send the related information to the BBU pool. The BBU pool then assigns the proper frequency Fig. 4. Different enhancements for C-RAN architectures in academia. for cell sites based on the received information, which can 1) Flexible C-RANs: The traditional C-RAN architecture effectively mitigate the interference thus boost the system is fully-centralized and fixed, which is not adaptive to the capacity. moveable traffic and the advanced software defined concept. 3) Other Enhanced Architectures: The traditional C-RAN As a result, it is urgent to improve the friable capability of C- is proposed to improve SE and EE, in which the control RANs. In [24], the novel concept of re-configurable fronthaul signal design is not optimized, and the interworking with the is proposed to flexibly support one-to-one and one-to-many existing cellular networks is not considered as well. Focusing logical mappings between BBUs and RRHs to perform proper on taking full advantages of both heterogenous networks and transmission strategies. Specifically, for areas with static users C-RANs, the authors in [29] [30] presented an H-CRAN as the and high traffic load, one-to-one logical mapping between advanced wireless access network paradigm, in which a new a BBU and a RRH will be configured to allow fractional communication entity named as Node C (Node with cloud frequency reuse to meet the traffic demand. While for areas computing) is defined. When this entity is used to converge with mobile and low traffic, one-to-many mapping will be macro BSs (MBSs), micro BSs, and other heterogeneous configured to allow distributed antenna systems to deliver nodes, it can be seen as a convergence gateway. When it is consistent coverage for mobile users and save the energy used to manage RRHs, it acts as a BBU pool. Meanwhile, consumption of the BBU pool. In [25], the challenges of four enabling techniques, including the advanced spatial signal re-configurable fronthaul, also named as software defined processing, cooperative radio resource management, network fronthaul (SDF), is further discussed, in terms of latency, function virtualization, and self-organizing network, are pro- communication protocol, and so on. In addition, the authors in posed to fulfill the potential of H-CRANs. Although several IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 7

TABLE III solutions have been proposed to alleviate the burden over the THE PERFORMANCESOF DIFFERENT FRONTHAUL TECHNOLOGIES fronthaul, the fronthaul in H-CRANs still suffers from the explosive traffic. To save fronthaul resource consumption and Fronthaul Tech- One-way Per-hop Latency Throughput nology Latency reduce delay, an F-RAN architecture is designed in [31], in NG-PON 2.5 µs 5 µs/km Over 10 Gbps which a large number of collaboration radio signal processing GE-PON 10-30 ms 1 ms 10 Mbps-10 Gbps and cooperative radio resource management functions can be EPON 5-10 ms 1 ms 10 Mbps-1 executed in fog computing based access points or by “smart” Gbps Digital 5-35 ms 5-35 ms 10 Mbps-100 UEs, and some packets can be directly delivered by the edge Subscriber Line Mbps devices through their limited cache. Access Cable 25-35 ms 25-35 ms 10 Mbps-100 Mbps Wireless 5-10 ms 5 ms 50 Mbps-1 D. C-RANs towards 5G Communication Gbps (200 MHz - 6 It is envisioned that 5G will bring a 1000x increase in terms GHz) Microwave < 1 ms 200 µs 100 Mbps-1 of area capacity compared with 4G, achieve a peak rate in Gbps the range of tens of Gbps, support a roundtrip latency of Millimeter wave < 1 ms 200 µs 500 Mbps-2 about 1 ms as well as connections for a trillion of devices, radio Gbps and guarantee ultra reliability [32], [33]. In [13], the field trials conducted by China Mobile have verified the throughput gain brought by C-RANs based on an uplink LTE model, In C-RANs, the fronthaul links carry information about the reaching up to near 300%. Through dense RRHs in C-RANs, baseband signals in the form of quantized I/Q samples, and massive connections are efficiently supported, and it is not the large bit rates produced by the quantized I/Q signals con- hard to provide good service for trillion of devices if the tradicts the limited capacity of the practical fronthaul solutions density of RRHs is sufficiently high. Although a big gap like the dark fiber solution [13], which has a significant impact is still observed compared to 5G requirements, the result on the LSCP gains [16]. In the uplink, for instance, RRHs has shown the potential advantages of C-RANs. Meanwhile, need to sample, quantize, and then forward the received I/Q different advanced techniques can be involved in C-RANs to signals to the BBU pool. With densely deployed RRHs, the further improve the spectrum efficiency, including CR, massive fronthaul traffic generated from a single UE with several MHz MIMO, and full duplex radio. Moreover, mmWave spectrum bandwidth could be easily scaled up to multiple Gbps. In can be used to provide much available bandwidth. As a result, practice, a commercial fiber link with tens of Gbps capacity it can be anticipated that C-RANs will be qualified for meeting could thus be easily overwhelmed even under moderate mobile the high SE and EE in 5G. traffic. Therefore, to tackle this problem, one approach is to The traditional C-RAN architectures should be enhanced utilize the partial centralization structure though substantial to meet low latency and high reliability requirements. For signal processing capabilities are required on RRHs. The other example, the fully centralized C-RAN architecture proposed alternative is to adopt advanced techniques to optimize the per- in [3] puts all functions of the air interface at the BBU pool. formance under a fully centralized structure with constrained Hence, the centralized C-RAN architecture is liable to cause fronthaul. As it simplifies the functions and capabilities of long latency, and decrease the reliability when the fronthaul is RRHs, the latter solution is the focus of current research constrained. Thus, decentralization is an alternative. Recently, community, and one of the corresponding key techniques is the in academia, with the integration of remote clouds and dis- fronthaul compression [16]. The aim of fronthaul compression tributed cloudlets, fiber-wireless access networks can improve is to compress the I/Q rate across multiple RRHs for all cloud accessibility with low latency and high reliability [34]. served UEs adaptive to the available fronthaul capacity, which Meanwhile in industry, Nokia has proposed to change the is critical to alleviating the impact of fronthaul capacity role of mobile BSs by endowing them with some IT-based constraints. capabilities like localized processing and content storage to Generally speaking, current fronthaul compression tech- implement distributed radio access networks (D-RANs) [35]. niques can be classified into the quantization based com- pression, compressive sensing (CS) based compression, and spatial filtering. The first one mainly includes point-to-point III. FRONTHAUL COMPRESSION compression and distributed source coding in uplink C-RANs In recent years, next-generation passive optical network as well as point-to-point compression and joint compression (NG-PON) has been considered one of the most promising op- in downlink C-RANs. While the latter two are mainly used in tical access technologies, which is envisioned to support a high uplink C-RANs. For the quantization based compression, the data rate (over 10 Gbps). To save the building cost, the popular compression is achieved by quantizing original signals, and and industrial optical access technologies are Ethernet PON the variance of the quantization noise, often assumed to be (EPON) and Gigabit EPON (GEPON), which take advantage zero-mean complex Gaussian, should be optimized. For the of inexpensive and ubiquitous Ethernet equipment, and offer CS based compression and spatial filtering, the compression the transmission speed up to 1 Gbps and 10 Gbps, respectively is performed by multiplying received signals with a local [36]. The intuition of the performances of different fronthaul compression matrix at each RRH, which can reduce the technologies has been illustrated in Table III. dimension of the signals to be transmitted to the BBU pool. IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 8

8VHU6LJQDO &DSDFLW\ distributed Wyner-Ziv coding and single-user compression.  55+ OLPLWHG %%83RRO Different from [38], in which the quantization noise level )URQWKDXO

&RPSUHVVLRQ '

H optimization problem is solved on a per-RRH basis, [40]

F

'

R

H

  

P

F

S

  

R jointly optimized the quantization noise levels of all RRHs.

U

G

  

H

L

V

Q V

J It is shown that setting the quantization noise levels to be L

 55+ R Q &RPSUHVVLRQ proportional to the background noise levels is near optimal for sum-rate maximization when the signal-to-quantization-noise- ratio (SQNR) is sufficiently high. Furthermore, it is concluded Fig. 5. The uplink compression model in C-RANs. that the gap between the sum rate achieved by Wyner-Ziv coding based scheme and a cut-set like sum capacity upper bound is bounded by a constant when quantization noise levels Note that the distributed source coding and joint compression are set to be proportional to the background noise levels. are often called by network-aware compression technique. The With single-user compression, a similar result exists under a core idea of distributed source coding is to enable the BBU diagonally dominant channel condition. pool to leverage the correlation among the signals received by Although the fronthaul loading can be effectively reduced neighboring RRHs, while CS is to leverage the sparsity of the by proposals in [39], [40], the uplink signal sparsity structure is UEs’ signals. Hence, such techniques are both available for ignored. In [41], by exploiting the signal sparsity in uplink C- uplink and are not directly applied in downlink. RANs , the distributed CS and recovery techniques are used to compress the fronthaul loading distributively. The performance of the proposed end-to-end signal recovery algorithm is ana- A. Uplink Compression lyzed, and it is shown that the restricted isometry property In the uplink, each RRH compresses its received baseband can still be satisfied by the aggregate measurement matrix signal and forwards the compressed data to the central BBU in C-RANs containing both distributed fronthaul compression pool as a soft relay through a limited-capacity fronthaul link and multi-access fading. While in [42], to realize a better as shown in Fig. 5. To achieve uplink fronthaul compression, design of the cloud decoder, joint decompression and decoding there are four main approaches: point-to-point compression are performed at the BBU pool. The sum-rate maximization [37], distributed source coding, CS based compression, and problem is a difference of convex problem. By numerical re- spatial filtering. When the low implementation complexity is sults, the advantage of the proposed algorithm is demonstrated, preferred, the point-to-point compression and spatial filtering compared to the conventional approach based on separate are recommended. When signals received at different RRHs decompression and decoding as in [38]. are closely correlated, the distributed source coding should be It is worth noting that in the presence of time-varying used. Considering the sparsity of the uplink C-RAN signals, channels, the overhead of communicating the CSIs from RRHs CS based compression is an another alternative. At the BBU to the BBU pool can becomes sufficiently large. Thus it is pool, the joint decoding operation based on all quantization essential to study the transfer of CSIs and data from RRHs values from all RRHs is executed. The state of the art of uplink to the BBU pool. In [43], the research on joint signal and fronthaul compression is briefly summarized in Table IV. CSI compression is conducted. Motivated by results in [44] In [38], it is observed that the imperfect knowledge of the related to the separation of estimation and compression, an joint statistics of the signals received at RRHs can lead to a estimate-compress-forward (ECF) approach is deployed, in performance degradation of distributed source coding. To deal which CSIs are first estimated at each RRH, and then the with this problem, a robust distributed compression scheme estimated CSIs are compressed for the transmission to the is proposed. Moreover, the network EE is taken into ac- BBU pool. Based on ECF approach, various strategies for the count by addressing RRH selection, in which an optimization separate or joint compression of the estimated CSIs and the problem aiming at the sum rate maximization is formulated received data signal are proposed and analyzed. When there by introducing a sparsity-inducing term to jointly consider are multiple RRHs, the proposed strategies can be combined the compression and RRH selection. Simulation results show with distributed source coding to leverage the received signal the effectiveness of the robust scheme on compensating the correlation across RRHs. It is observed that the ECF approach performance loss due to the imperfect statistical information. outperforms the compress-forward-estimate approach as used Still tackling the issue of robustness in distributed compres- in [45]. sion, the authors of [39] investigated the layered transmission In [46], the authors considered the uplink of a C-RAN with and compression strategies. Under competitive robustness and multi-antenna RRHs each connected to the BBU pool through fronthaul capacity constraints, the compression is formulated a finite-capacity fronthaul link, and proposed a novel spatial- as the minimization of the transmit power. It is shown that compression-and-forward (SCF) scheme. Unlink literatures the proposed layered strategies greatly outperform single-layer [38], [40] adopting complex quantization schemes, a simple strategies, and it is the layered transmission not the layered linear filter is used to compress the correlated signals received compression that contributes the main performance gains. by all the antennas at each RRH. An optimization problem is Authors in [40] formulated the problem of optimizing the formulated, aiming at maximizing the minimum SINR of all quantization noise levels to maximize weighted sum rate the users by jointly optimizing users’ power allocation, RRHs’ under a sum fronthaul capacity constraint by involving the spatial filter design and quantization bits allocation, and BBU’ IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 9

8VHUVLJQDO &DSDFLW\ %%83RRO the separate design of precoding and compression. However, OLPLWHG 55+  8VHU0HVVDJHV )URQWKDXO

(QFRGLQJ &RPSUHVVLRQ the work [49] has its own limitations that the optimization of

&

R

3 P

U the pre-coding and the joint compression is only within a given

   

S

H

U

F

   

H

R

    V

G 8VHUVLJQDO V

L cluster of RRHs without considering the effect of inter-cluster

Q

L

R J Q 55+  interference. When applied to scenarios of multiple mutually (QFRGLQJ &RPSUHVVLRQ interfering clusters, inter-cluster interference could degrade the performance of its approach. Based on this observation, Fig. 6. The downlink compression model in C-RANs. authors in [50] investigated the joint design of pre-coding and fronthaul compression in a multi-cluster case, in which the problem of maximizing the weighted sum-rate across all s receive beamforming. clusters is formulated, and an iterative algorithm is proposed to Whereas in [47], the quantization strategy optimization is find the solution to the problem. The simulation results show handled for the positioning in C-RANs, and its goal is to that a significant gain can be reaped by the proposed scheme. minimize the worst-case localization error of a single-antenna radio transmitter under fronthaul capacity constraints. The Both in [49] and [50], the joint pre-coding is executed at Charnes-Cooper transformation and difference-of-convex pro- the BBU pool, which is classified into the pure compression gramming are utilized to design corresponding optimization strategy. The pure compression means that the precoded analog algorithms, and several conclusions are got by numerical signals are compressed and forwarded to the corresponding simulations. For example, it is observed that a larger fronthaul RRHs in downlink, which is akin to the compress-and-forward capacity brings an improved localization, and the localization relaying strategy. To implement the joint pre-coding, a pure has lower fronthaul rate requirements compared with data message-sharing strategy [51] can be used, in which the BBU communication. pool simply shares each user’s message with multiple RRHs Most of the aforementioned works only considered the over the fronthaul links. The pure message-sharing strategy scenario where RRHs are directly connected to the BBU pool. can be thought of as analogous to the decode-and-forward Instead, in [48], a general multi-hop fronthaul scenario is relaying strategy. Through combining advantages of both pure studied, in which each RRH may communicate with the BBU compression and pure message-sharing, a hybrid strategy pool through a set of intermediate RRHs. In this scenario, adaptive to the available fronthaul capacity is proposed in the multiplex-and-forward (MF) strategy may suffer from a [51], whose aim is to allocate some fronthaul capacity to send significant performance loss when RRHs are densely deployed. direct messages for some UEs (for whom RRHs are better To solve this problem, a decompress-process-and-recompress off receiving messages directly, instead of their contributions (DPR) strategy is proposed, in which each RRH decompresses in the compressed precoded signals) by the pure message- the received streams and performs linear in-network process- sharing strategy, and the remaining fronthaul capacity is used ing of the decompressed signals. For both MF and DPR to carry the compressed signal that combines the contributions strategies, the optimal design is handled to maximize the sum- from the rest of UEs by the pure compression strategy. The rate under fronthaul capacity constraints. By comparing the analysis and simulation results show that the hybrid strategy performance of MF and DPR strategies, the advantage of in- is more beneficial to the overall system performance in a network processing is highlighted. practical C-RAN with finite fronthaul capacity, compared with any separated compression strategy. B. Downlink Compression 2) Compression under Dynamic Channels: Different from Fig. 6 shows an example of the downlink compression in literatures [49]–[51] assuming static channels and ideal CSIs C-RANs. First, the BBU pool performs channel coding and at the BBU pool, the authors in [52] investigated the joint pre-coding for users’ message, and before transmission on fronthaul compression and precoding design with a block- the fronthaul, the pre-coded signals are compressed jointly or ergodic fading channel model in the downlink C-RAN with separately. Then RRHs decompress the received signals and a cluster of RRHs with multiple antennas serving multi- transmit them to the desired users. The state of the art of antenna users. The analysis for two types of BBU-RRH func- downlink fronthaul compression is briefly surveyed, and the tional splits at PHY are conducted, which corresponds to the main literatures are summarized in Table V. compression-after-precoding (CAP) and compression-before- 1) Compression under Static Channels: In [49], multi- precoding (CBP) strategies, respectively. For the first strategy, variate compression is leveraged to suppress the additive all the baseband processing is performed at the BBU pool, quantization noises at UEs. The joint pre-coding and com- while for the second strategy, RRHs perform channel encoding pression design problem to maximize the weighted sum-rate and precoding instead of the BBU pool which only separately is formulated under power and fronthaul capacity constraints, forwards the user message and the compressed precoding and a stationary point is reached by a proposed iterative matrices to each RRH. The ergodic capacity optimization for algorithm. Moreover, the practical implementation of the mul- both strategies is tackled by the proposed stochastic successive tivariate compression and the robust design of the pre-coding upper-bound minimization approach. The numerical results and compression are tackled as well. Numerical results have show that the optimal BBU-RRH functional split depends on demonstrated that the proposed joint precoding and compres- the interplay between the enhanced interference management sion strategy outperforms conventional approaches based on abilities of CAP and the lower fronthaul requirements of CBP. IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 10

TABLE IV SUMMARY OF LITERATURESRELATEDTO UPLINK FRONTHAUL COMPRESSION

Literature Compression System model Optimization Main contribution Method objective [38] Distributed Multiple multi-antenna RRHs and multi-antenna Maximize the Propose a robust distributed compression scheme source coding users scenario with per-RRH fronthaul capacity sum rate that can compensate the performance loss due to constraint the imperfect statistical information [40] Distributed Multiple single-antenna RRHs and single-antenna Maximize the Find near optimality of setting the quantization source coding users scenario with a sum fronthaul capacity weighted sum noise levels to be proportional to the background and point- constraint rate noise levels for sum-rate maximization under high to-point SQNR compression [41] CS based com- Multiple single-antenna RRHs and single-antenna Maximize the Exploit the signal sparsity in the uplink C-RAN pression users scenario without fronthaul capacity con- probability of and analyze the tradeoff relationship between the straint correct active C-RAN performance and the fronthaul loading user detection [48] Point-to-point Multiple multi-antenna RRHs and multi-antenna Maximize the Propose a DPR strategy and highlight the advan- compression users scenario with a multihop fronthaul topology sum rate tage of in-network processing [46] Spatial filtering Multiple multi-antenna RRHs and single-antenna Maximize Show that a multi-antenna C-RAN generally sig- users scenario with per-RRH fronthaul capacity the minimum nificantly outperforms both massive MIMO and constraint SINR of all a single-antenna C-RAN under practical fronthaul users rate constraints [47] Point-to-point Multiple single-antenna RRHs and a single- Minimize the Propose optimization algorithms based on Charnes- compression antenna user scenario with per-RRH fronthaul worst-case Cooper transformation and difference-of-convex capacity constraint localization programming error of the target

TABLE V SUMMARY OF LITERATURESRELATEDTO DOWNLINKFRONTHAUL COMPRESSION

Literature Compression System model Optimization Main contribution Method objective [49] Joint compres- A given cluster composed of multiple multi-antenna Maximize the Suppress the additive quantization noises at sion RRHs and multi-antenna users with per-RRH fronthaul weighted sum users by leveraging multivariate compression capacity constraint rate [50] Joint compres- Multiple mutually interfering clusters each of which Maximize the Joint design of pre-coding and fronthaul com- sion consists of multiple multi-antenna RRHs and multi- weighted sum pression in a multi-cluster case antenna users with per-RRH fronthaul capacity con- rate straint [51] Point-to-point Multiple single-antenna RRHs and single-antenna Maximize the Propose a hybrid scheme of pure compression compression users scenario with per-RRH fronthaul capacity con- weighted sum and pure message-sharing, achieving signifi- straint rate cant performance gains [52] Point-to-point Multiple multi-antenna RRHs and multi-antenna users Maximize the Joint fronthaul compression and precoding compression scenario with per-RRH fronthaul capacity constraint ergodic capac- design with a block-ergodic fading channel ity model, which yields insights into the optimal BBU-RRH functional split for C-RANs

C. Lessons Learned processing should be jointly optimized to alleviate the fron- thaul constraint. Meanwhile, the network-aware compression Due to the constrained fronthaul and decreasing the over- requires the full and ideal CSIs of both the access links head of transmitting I/Q signals, the fronthaul compression between RRHs and UEs and the wireless fronthaul links, is key for both uplink and downlink in C-RANs. Point-to- which are interesting open problems in uplink. point compression can reduce the bit rate of fronthaul through compression as applied separately on each fronthaul link, while the network-aware compression can provide a significant compression gains, which towards the network information- IV. LARGE-SCALE COLLABORATIVE PROCESSING theoretic optimal performance. In uplink, the network-aware compression techniques include the distributed source coding Without advanced interference coordination techniques in and spatial filtering. While in downlink, since the point-to- C-RANs, interference should be severe due to the dense point compression is suboptimal from a network information- deployment of RRHs, which significantly decreases SE and EE theoretic viewpoint, the joint multivariate compression is of C-RANs. To mitigate the interference and thus improve SE preferred. Comparing with downlink, the uplink compres- and EE, the centralized LSCP techniques should be executed sion is more challenging because I/Q signals from UEs are in the BBU pool. Currently, several approaches have been sparsely and massively distributed. Generally, the network- proposed to solve this challenge as shown in Fig. 7, includ- aware compression technique outperforms the separate design ing the joint precoding and fronthaul compression design to of precoding and compression. Compared with the separated alleviate the fronthaul capacity constraint, the compressive strategy [51], the hybrid of pure compression and message- CSIs acquisition and stochastic beamforming to deal with the sharing is more beneficial to the overall system performance difficulty in obtaining perfect CSIs, and the sparse precoding in a practical C-RAN with finite fronthaul capacity. In the design to address RRH selection. In this section, the related future, the network-aware compression and the sparse signal papers are summarized in Table VI. IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 11

TABLE VI SUMMARY OF LITERATURESRELATEDTO LARGE-SCALE COLLABORATIVE PROCESSING

Literature Classification Problem Description Challenge Solution [59] Precoding with Joint RRH selection and linear precoding High complexity for RRH selection with the guidance of the spar- perfect CSIs design for the downlink transmission in RRH selection sity pattern of the precoding matrix networks with multi-cell processing [49] Precoding with Joint precoding and fronthaul compres- The control of the effect Multivariate compression perfect CSIs sion design for downlink C-RANs of additive quantization noises at the MSs [50] Precoding with Inter-cluster design of precoding and The existence of mutu- Inter-cluster optimization of precoding and perfect CSIs fronthaul compression ally interfering clusters joint compression by a proposed iterative al- gorithm [74] Precoding with Beamforming design for a downlink Huge overhead to obtain Compressive CSIs acquisition and stochastic imperfect CSIs Cloud-RAN full CSIs beamforming [76] Precoding with MMSE precoding algorithm for a down- Channel estimation errors Bayesian philosophy for reducing the effects imperfect CSIs link Cloud-RAN and low-complexity algo- of channel estimation errors and alternating rithm design direction method of multipliers for decreasing computational complexity [77] Precoding with Beamforming for multicast green Cloud- An infinite number of The adoption of convex optimization tech- imperfect CSIs RAN the non-convex quadratic nique based on PhaseLift, semidefinite relax- QoS constraints ation and S-lemma [78] Sparse precod- Beamforming for limited-fronthaul net- The non-convexity of the Iteratively relaxing the l0-norm as a weighted ing work MIMO system problem of finding the l1-norm optimal tradeoff between the total transmit power and the sum fronthaul ca- pacity [79] Sparse precod- Joint beamforming and clustering design Per RRH fronthaul con- Iteratively approximating the per RRH fron- ing for a downlink network MIMO system straints with l0-norm thaul constraints by a reweighted l1-norm technique [60] Sparse precod- Joint RRH selection and power mini- High complexity for Group sparse beamforming method based on ing mization beamforming for a Cloud-RAN solving the formulated the group-sparsity of beamformers induced by problem the weighted l1/l2-norm minimization

precoding matrix gives some information on RRH selection. Virtualized BBU Pool For previous RRH selection schemes, RRHs with shorter

Joint Precoding Compressive CSI and Fronthaul Acquisition and Sparse Precoding distances from mobile stations or with better channel con- Compression Stochastic Design Design Beamforming ditions are chose. Differently, the algorithm proposed in [59] selects the cooperating RRHs based on the regularized convex

&DSDFLW\OLPLWHG IURQWKDXO optimization. Numerical results suggest that the proposed

 algorithm can achieve a significant performance gain with a &OXVWHU   'HFRPSUHVVLRQ   low computational complexity. 2)) 21 To further overcome the indispensable limitation of fron- 8( 55+ thaul capacity, the authors in [49] studied the joint design of precoding and compression strategies for the downlink C-RAN with limited fronthaul capacity. DPC and linear precoding Fig. 7. Different precoding designs in C-RANs. are compared, and the numerical results show that DPC has better performance only in the regime of intermediate transmit power due to the limited-capacity fronthaul links, which A. Precoding with Perfect CSIs suggests that the compression has bigger impact on the system Several literatures have been published for the precoding performance rather than the precoding methods. However, design recently, most of which are presented to target the the paper mainly focuses on the optimization of precoding mitigation of interference and the improvements of SE and coefficient design within a single cluster of RRHs, and the EE performances. Dirty paper coding (DPC) [53] [54] achieves corresponding performance gain is marginal if considering capacity region, however, the complexity is too high especially interference from other clusters. To this end, the authors of for large-scale C-RANs. Therefore, there is few published [50] considered a weighted sum rate maximization problem works considering DPC for the large-scale C-RANs. On the subject to fronthaul capacity and per RRH power constraint other hand, the linear precoding is regarded as a low complex with the aim of optimizing the precoding and quantization alternative [55]–[58]. The linear precoding techniques mainly covariance matrices across all clusters. An iterative algorithm utilize CSIs in a simple and efficient way to manage interfer- based on the majorization minimization approach is utilized to ence across users and among the data streams for the same achieve a stationary point, in which the inter-cluster design, user. inter-cluster time-division multiple access and intra-cluster 1) Precoding Design with Limited Fronthaul Capacity: design algorithms are proposed to enable the achievable rates Considering the fronthaul capacity is limited in practice and to approach the near-optimal configuration. It is learned the the load increases with the number of cooperating RRHs, a proposed precoding based on multivariate compression and joint RRH selection and linear precoding design algorithm inter-cluster design significantly outperforms the intra-cluster is investigated in [59], in which the sparsity pattern of the approach in [49]. IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 12

2) Precoding Design for EE Maximization: Another aspect CSIs to guarantee the QoS requirements. It is revealed that the is to maximize the EE performance. Due to the huge power proposed SCB scheme can significantly reduce the overhead consumption caused by a large number of RRHs and fronthaul while providing performance close to that with full CSIs. links, the power consumption has a significant impact on However, it is noted that the required number of samples EE [60]. Considering the fact that EE is one of the major increases rapidly with the size of networks. Given the fact objectives in the future cellular networks, allowing some RRHs that [74] only provides some feasible solutions without any to enter sleep mode is beneficial when the traffic load is optimality guarantee, authors in [75] proposed a novel stochas- not huge in the corresponding area [61]. For the power tic difference-of-convex programming algorithm to solve this consumption of the fronthaul links, it is determined by the SCB problem, and it can reach the globally optimal solution set of active RRHs, and the transmit power consumption of if the problem is convex and reach a locally optimal solution the active RRHs can be minimized through the coordinated if the problem is non-convex. It is learned that the CSIs beamforming. As a result, a joint RRH selection and power acquisition overhead can be reduced by about 40%. minimization problem is built in [60], in which the partial 3) Robust Precoding Design: In [76], a robust low- precoding design is proposed by applying optimization with complexity minimum mean square error precoding design group sparsity induced norm. The sparsity patterns in signal problem is investigated, aiming at minimizing the total average processing have been exploited for a high efficient system mean square error with per-antenna power constraint. Here, design [62]. The l1-norm regularization has been successfully imperfect CSIs is concerned due to the Gaussian distributed applied in compressed sensing [63], [64]. On the other hand, channel estimation errors. The alternating direction method the mixed l1/l2 -norm [65] and l1/l∞-norm [66] are most of multipliers (ADMM) is employed to decouple the op- commonly used to induce the group sparsity because of their timization objective and the constraints, greatly decreasing analytical and computational convenience. the complexity of the proposed robust algorithm, especially for the large-scale C-RANs. Simulation results show that the proposed ADMM-based robust design can improve the system B. Precoding with Imperfect CSIs performance. 1) Precoding with Statistical CSIs: Considering the BBU 4) Precoding Design for Energy Minimization: The authors pool can typically support hundreds of RRHs, it is apparent in [77] focused on the energy minimization problem for that the acquisition of full CSIs is critical to the optimal design the multicast C-RAN under imperfect CSIs, which includes of precoding in C-RANs. Note that many works commonly an infinite number of non-convex quadratic QoS constraints. assumed the BBU obtains the full CSIs through pilot training, A computationally efficient and robust algorithm with three however, obtaining full CSIs is challenging because a lot stages is proposed. In the first stage, a novel quadratic varia- of parameters are involved, leading to significant estimation tional formulation of the weighted mixed l1/l2-norm is used errors, quantization errors, and long feedback delays [67]. to induce the group-sparsity structure for the robust multicast Fortunately, partial CSIs can be more accurately obtained beamformer for the guidance of RRH selection. The perturbed through long term channel feedback. Precoding based on alternating optimization algorithm is then considered to solve the statistical CSIs has been well studied for single-user the resultant non-convex group-sparsity inducing optimization multiple input multiple output (SU-MIMO) [68], [69] and problem. In the second stage, a PhaseLift method is introduced then extended to multiuser MIMO (MU-MIMO) [70], [71]. to handle the feasibility problems in the RRH selection proce- A scheme firstly proposed in [72] considers a transmission dure, and the semidefinite relaxation technique is adopted to scheme that is proved to achieve a larger degree of free- design the beamformer in the final stage. dom than one with completely outdated CSIs. This scheme alongside precoding is used to improve the achievable rate in [73]. On this account, the statistical CSIs in conjunction C. Sparse Precoding with this transmission scheme is utilized in [67] to develop In C-RANs, the power consumption of fronthaul has a both transmission and precoding schemes that provide further significant impact on EE. Hence, allowing the fronthaul links system performance improvement. and the corresponding RRHs to support the sleep mode are es- 2) Stochastic Coordinated Beamforming: To overcome the sential to reduce the power consumption for C-RANs. It would curse of dimensionality for C-RANs and provide an CSIs be feasible to switch off some RRHs while still maintaining the overhead reduction, the authors in [74] proposed a novel CSIs QoS requirements. Group sparsity is required rather than the acquisition scheme called compressive CSIs acquisition, which individual sparsity as used in compressed sensing [63] [64], can obtain the instantaneous CSIs of a subset of channel similar to [65] [66]. In [60], a mixed l1/lp-norm is adopted to links and the statistical CSIs of the others. Through this induce sparsity in large-scale C-RANs, however, the approach scheme, more overhead can be saved due to the reduced in [65] [66] simply considered un-weighted mixed l1/lp-norms channel compression dimension in large-scale C-RANs. With to induce the group-sparsity, in which no prior information of these mixed CSIs, a coordinated beamforming is conducted the unknown signal is assumed other than the fact that it is to minimize the total transmit power while satisfying the QoS sufficiently sparse. requirement for all UEs. The authors proposed a stochastic In [78], considering the network resources including the coordinated beamforming (SCB) framework based on the fronthaul capacities and the transmit powers at RRHs, an chance-constrained programming by exploiting the statistical optimal joint clustering and beamforming design problem is IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 13 addressed, in which each user dynamically forms a sparse D. Lessons Learned network-wide beamforming vector whose non-zero entries One of the key advantages in C-RANs is to provide a correspond to the serving RRHs. By adopting a compressive large-scale centralized signal processing across multiple RRHs sensing approach of using reweighted l1-norm to approximate and UEs in both uplink and downlink. The feasibility and the l0-norm, the authors turn the original non-convex problem performance gains of LSCP depends strictly on CSIs. If into a series of convex weighted power minimization problem, the capacity of fronthaul is unlimited, the well-known linear further solved by a low-complex duality approach. Simulation beamforming and successive interference cancellation can be results show that the proposed algorithm can achieve a better directly applied in uplink, and the minimum mean-square error tradeoff between the total transmit power and the sum fron- or zero-forcing beamforming strategies can be used to achieve thaul capacity than the existing methods in the high SINR a good performance gain. On the other hand, a broadcast regime. channel is formed and the linear precoding technique can As a further enhancement of [78], which mainly consid- be directly used in downlink C-RANs when the capacity of ers the sum power and sum fronthaul capacity, the authors fornthaul is unlimited. of [79] [80] focus on network utility maximization under Since the fronthaul is often capacity constrained, LSCP per-RRH transmit power and the per-RRH fronthaul capac- and compression process should be jointly designed, and the ity constraints. Among the family of utility functions, the core challenge for LSCP in C-RANs is to design the efficient weighted sum rate (WSR) has been widely applied to the algorithms when CSIs are non-ideal and the fronthaul is network control and optimization problems. Motivated by the constrained. As an initial LSCP strategy in uplink C-RANs, idea in compressive sensing literatures, non-convex l0-norm fixed-rate scaler uniform quantization is preferred when the in the fronthaul constraint is approximated by the convex practical LSCP is designed. In downnlink, the zero-forcing reweighted l1-norm. Furthermore, the WSR problem can be beamforming strategy with scalar fixed quantization can be reformulated as an equivalent weighted minimum mean square first used. Furthermroe, substantial overhead to obtain full error (WMMSE) minimization problem in order to reach a lo- CSIs for precoding design can be reduced by compressive cal optimum solution. To reduce the computational complexity CSIs acquisition and stochastic beamforming [74]. With the of the proposed algorithm, it is proposed to iteratively remove guidance of the sparsity pattern of the precoding matrix, the RRHs with negligible transmit power out of the candidate complexity of RRH selection can be significantly decreased cluster for each UE, and remove those UEs with negligible [59]. Finally, per RRH fronthaul constraints with l0-norm can rates out of the scheduling pool. be handled by iteratively approximating the per RRH fronthaul On the basis of [79], [80] considered the static clustering constraints through a reweighted l1-norm technique [79]. which fixes the RRH cluster for each UE and jointly optimized the scheduling and beamforming vectors. Two static clustering V. CHANNEL ESTIMATION AND TRAINING DESIGN formation algorithms are proposed, namely the maximum In C-RANs, the lack of full CSIs presents the main hurdle loading based static clustering and the biased signal strength to the feasible implementation of C-RANs since an accurate based static clustering, which both effectively take the traffic acquisition of instantaneous CSIs is essential for most of load of each RRH and the channel condition of each UE into advanced technologies, e.g., interference coordination and account. A novel point in [80] is to incorporate the per-RRH coherent reception. The assumption that the full CSIs are fronthaul constraint into the optimization framework. With known at the BBU pool is also not practical due to the time- explicit per-RRH fronthaul constraints, it is shown that the varying nature of the radio channel. Consequently, developing fronthaul resources can be efficiently utilized and the network accurate and efficient channel estimation techniques is vital utility can be significantly improved. to achieve the performance improvement of C-RANs [84]. A framework is proposed to design a green C-RAN in [60], In this section, an overview of channel estimation techniques which is formulated as a joint RRH selection and power min- proposed for C-RANs is provided to give useful guidelines imization beamforming problem. Considering that the power on the new requirements when employing the conventional consumption of fronthaul is determined by the set of active channel estimation approaches in C-RANs. Moreover, some RRHs, a group sparse beamforming method is proposed by in- challenging issues about channel estimation in the fronthaul ducing the group-sparsity of beamformers using the weighted constrained C-RANs are presented as the meaningful research l1/l2-norm. The proposed problem is a mixed-integer non- topics in the future. linear programming problem, which is NP-hard. To solve this Channel estimation as a fundamental problem in wireless problem, two group sparse beamforming (GSBF) formulation communication systems has been extensively investigated in algorithms with different complexities are proposed, namely, the past decades [85]. In general, channel estimation can be bi-section GSBF and iterative GSBF. It is demonstrated that basically categorized into two kinds of strategies, i.e., training- the GSBF framework is effective to provide near-optimal based and non-training-based channel estimation. The training- solutions to this problem. The proposed bi-section GSBF based channel estimation is the most common method in the algorithm is proved to be a better option for a large-scale CSIs acquisition of wireless communication systems due to C-RAN due to its low complexity, while the iterative GSBF its flexibility and low complexity. By periodically inserting algorithm can be applied to provide better performances in a the training sequences into data frames in time or frequency medium-size C-RAN. domain, the receiver can easily estimate the CSIs of the IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 14

radio channels by a prior knowledge of the training sequences ;YKX+W[OVSKTZ BBU Pool [86]. However, the use of the training sequences degrades 8KSUZK8GJOU.KGJ

JGZGY_SHUR the system performance in terms of SE and EE as additional ,XUTZNG[R2OTQY ')2ZXGOTOTM spectrum resources are demanded for the training transmission. 8GJOU'IIKYY2OTQY Therefore, the non-training-based channel estimation is intro- =,2ZXGOTOTM duced as an alternative way to obtain the CSIs by adequately using the intrinsic characteristics of data signal itself instead 88. of pilot signal [87]. Despite the high SE, there still exits some restrictions of non-training-based channel estimation, e.g., slow convergence speed, high computation complexity, and sensitivity to noise, which limit the implementation in Fig. 8. Training scheme in C-RANs. practice. In the next, as shown in Table VII, the training-based channel estimation in C-RANs is mainly discussed as well as the applications of non-training-based channel estimation in judging the discriminant of the polynomial. To explore the C-RANs. auto-correlation nature of time-selective fading channel, a correlation-based iterative channel estimation algorithm in [90] can further improve the estimation performance in MARNs by A. Superimposed Training Based Channel Estimation in C- observing the instantaneous training signals. It is proved that RANs the iterative channel estimation algorithm is convergent and The data transmission in C-RANs is performed by means of can provide better accuracy than the ML method. two communication links, the first part by the radio access link (ACL) between the UE and the RRH, and the remaining part B. Segment-Training Channel Estimation by the fronthaul link connecting the RRH and the BBU pool. In the superimposed training scheme, the data transmission The fronthaul links are wire/wireless, and the wireless fron- rate is still not high and the accuracy of channel estimation thaul links (WFLs) are mainly discussed herein due to its low is not efficient due to the fact that training sequences and expenditures and flexible deployment for RRHs. To achieve data symbols have to share the transmitted power. On the the optimal system design in C-RANs, the individual CSIs of other hand, the performance of channel estimation is greatly these two links should be obtained at the BBU pool as required influenced by the unknown data symbols when extracting by certain technologies, e.g., beamforming design [88]. Under the training sequences from the compounded signals, and the circumstances, the traditional training design is inefficient it is difficult to apply in practice due to the sensitivity to for fulfilling the requirements due to the high training overhead distortion. The segment training scheme in [91] can be applied and the significant performance degradation resulting from to acquire the individual CSIs of the ACLs and the WFLs in the CSIs feedback during the individual CSIs acquisition. To C-RANs, which is an effective way to improve the estimation tackle this problem, the superimposed training scheme in [89] accuracy. In the segment training scheme, two consecutive is a promising solution to the channel estimation in C-RANs, segments as illustrated in Fig. 8 are implemented for the CSIs in which the RRH superimposes its own training sequence on acquisition of ACLs and WFLs, respectively. At the beginning, the received one such that the individual CSIs can be estimated the CSIs of the WFLs can be obtained at the BBU pool by at the BBU pool. In this way, a better tradeoff between the using the first segment training. Then, based on the estimated channel estimation accuracy and the training overhead can CSIs of the WFLs, the CSIs of the ACLs is readily acquired be achieved in the superimposed training scheme with no by employing the training-based channel estimation method. occupation of additional training resources compared with the Generally, the segment training scheme can achieve higher traditional training scheme. estimation accuracy as the power originally shared by the The discussion on the superimposed training based channel training and data is allocated to the training alone. The segment estimation can first start with channel estimation in multiple training based channel estimation can be recognized as an access relay networks (MARNs) [90], which can be extended alternative for the acquisition of CSIs in C-RANs when a high to C-RANs. As shown in Fig. 8, the superimposed training estimation accuracy is required. scheme in [90] considers two phase. In the superimposed 1) Balance between Performance and Overhead: It should training scheme, the sources transmit the training matrix T in be noted that straightforward implementation of the segment the first phase and each RRH superimposes a training sequence training scheme in C-RANs would degrade SE since the tr over the received signal in the second phase. By using these number of symbols assigned for the training is sufficiently two training signals, the destination node can separately obtain large. A superimposed-segment (SS) training scheme in [92] the CSIs of the composite source-RRH-destination and the can be regarded as a good way to achieve the balance between individual RRH-destination links. Based on the superimposed the data rate and training overhead in C-RANs. In the SS training scheme, the channel estimation in uplink network training scheme, the superimposed training ˜tr is implemented coded MARNs is investigated by employing the maximum for the ACLs to reduce the training overhead, while another likelihood (ML) method. Furthermore, it is pointed out that individual segment is assigned for the training ts of the the optimal training structures of tr and T are orthogonal. WFLs to avoid the data rate reduction, which is shown in The optimal power allocation strategy can be derived by Fig. 8. Due to the time-varying characteristics of the ACLs, IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 15

TABLE VII SUMMARY OF CHANNEL ESTIMATION TECHNIQUESIN C-RANS

Literature Techniques Approaches Feature [86] Training-based channel estimation Training sequences embedded into data Easy deployment, high accuracy, technical frames and CSIs acquisition based on a prior maturity and low spectral efficiency knowledge of training [89], [90], [92] Superimposed training based channel Training sequences superimposed on the data High resource utilization, poor estimation per- estimation symbols formance and low received signal-to-noise ra- tio [91], [93] Segment training based channel esti- Two consecutive segments to obtain the indi- High estimation accuracy, low spectral effi- mation vidual CSIs ciency and data transmission rate guarantee [87] Blind/semi-blind channel estimation Channel estimated without training by using High spectral efficiency, long observation time the statistical and other properties of the noise and high computational complexity and data the complex-exponential basis-expansion-model is adopted to RANs if merely relying on the data samples to estimate the model the ACLs. To obtain the individual CSIs of ACLs CSIs. Still, a small number of training is needed to obtain and WFLs, the maximum a posteriori probability channel an initial estimate. Semi-blind channel estimation as a hybrid estimation algorithm can be applied. By iteratively calculating of blind and training based approaches is considered as a the estimates of ACLs and WFLs, the channel estimation pragmatic compromise to take their strong points to offset the accuracy can be further improved. Moreover, the iterative weakness. By incorporating both the training signal and the channel estimation algorithm is convergent, and the channel data samples, the semi-blind channel estimation can achieve estimation accuracy is higher than the ML method. higher accuracy than purely using the training-based or blind 2) Estimation in Large-scale C-RANs: In the large-scale C- estimation. Moreover, the study of the semi-blind channel RANs, the channel estimation is still challenging since there estimation in C-RANs is in its initial phase, and there are are lots of channel parameters to be estimated simultaneously. only tentative attempts at the related research area. As a result, the computational complexity is another important Inspired by the work in [95], the semi-blind channel esti- issue for the channel estimation in C-RANs. The conventional mation is employed in C-RANs to achieve a good accurate linear minimum-mean-square-error (LMMSE) estimator has a estimation without sacrificing the bandwidth efficiency. The cubic complexity in the dimension of the covariance matrices least-squared estimation is used to obtain an initial estimate, due to the matrix inversion, and it becomes awfully large to be and the quasi-Newton method is applied to further improve unaffordable in the large-scale C-RANs. For the complexity of the estimation accuracy based on the maximum likelihood MMSE estimator, [93] has successfully reduce it to a square principle. By treating data symbols as Gaussian-distributed complexity by applying the matrix polynomial expansion (PE). nuisance parameters, the joint likelihood function of both the PE is a well-known technique to reduce the complexity of training and data can be derived. The maximization problem large-dimensional matrix inversions by approximating a matrix of the joint likelihood function is actually a nonlinear problem, function by an L-degree matrix polynomial. Inspired by the and an iterative quasi-Newton method can be applied to obtain low computational complexity estimator, the same method can the CSIs in C-RANs with a limited number of data symbols. be applied to the channel estimation in C-RANs with segment In this way, the semi-blind channel estimation can achieve training scheme. First, the LMMSE estimator is employed to substantial improvements over the conventional training-based obtain the CSIs of the ACLs and WFLs. When implementing method in terms of estimation accuracy. the LMMSE estimator, the matrix inversion of the covariance matrix of the received training can be replaced by the PE. In this way, the computational complexity of the segment training D. Lessons Learned based channel estimation is significantly reduced. Compared with the traditional training design scheme, a good tradeoff between the channel estimation accuracy and C. Non-Training-Based Channel Estimation in C-RANs the training overhead can be achieved in the superimposed The training-based channel estimation has demonstrated its training scheme for C-RANs. Meanwhile, the non-training- effectiveness to acquire a reliable estimate of the CSIs in based channel estimation can achieve high SE but meanwhile C-RANs. However, the training overhead of training-based with high computing complexity. Accurately, by incorporating channel estimation is relatively significant, and the transmis- both the training signal and the data samples, the semi-blind sion bandwidth efficiency is decreased in turn. The training is channel estimation can achieve higher accuracy than the pure undesirable for some special scenarios, and the training-based training-based or blind estimation schemes. channel estimation has not exploited the intrinsic properties of Furthermore, since the constrained fronthaul is one of the data symbols. Thus, the non-training-based channel estimation major practical hurdles for C-RANs, the channel estimation can be employed to resolve the problem of spectrum resource by taking fronthaul loading into consideration is an important scarcity [94]. However, a large number of data symbols is and challenging issue, which needs to be urgently tackled in required to improve the channel estimation accuracy blindly. the future. The channel estimation errors should be taken into The inherent ambiguity, e.g., phase ambiguity, also blocks consideration and the corresponding influence is urgent to be the implementation of the blind channel estimation in C- exploited. IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 16

VI. COOPERATIVE RADIO RESOURCE ALLOCATION

To maximize SE and EE in C-RANs, the multi-dimensional Virtualized BBU Pool radio resources should be optimized, which often leads to non- convex problems. When the fast adaptation to users’ traffic is &DSDFLW\  required, cross-layer optimization needs to be tackled. In this OLPLWHG section, selected published works , related to the static CRRA IURQWKDXO   without QSI and the dynamic CRRA with queue-awareness,  are discussed and summarized. 55+ 8( DVVRFLDWLRQ 8(  VFKHGXOLQJ

3RZHU A. Static CRRA without QSI $OORFDWLRQ 55+FOXVWHU 55+ IRUPDWLRQ For the static CRRA without QSI, the classic methods like Lagrange dual method and WMMSE method can be directly adopted. In addition, to alleviate the transmission burden on Fig. 9. Multi-dimensional radio resource allocation in C-RANs. the fronthaul, the game model based optimization approaches can be used as well. 1) Classic Non-convex Optimization: The performances of can be equivalently transformed into a convex optimization C-RANs, such as SE and EE, are mainly determined by the problem, i.e., its dual problem. multi-dimensional CRRA, including the power allocation or Secondly, WMMSE has been widely applied to design the precoding design, user scheduling, RRH cluster formation, and centralized power allocation in C-RANs [80], [101], [103], so on, as shown in Fig. 9. To achieve the global optimal so- [104]. Due to the potential intra-tier interference and limited lutions to the static CRRA problems without QSI, centralized cluster size, the expression of SINRs is no longer convex methods are effective because the multi-dimensional CRRA with regard to the entries of the precoding matrix, causing the which influences the system performance can be monitored by non-convexity of the whole problem. Fortunately, WMMSE the BBU and optimized jointly. These centralized method have has been proved to be effective in transforming this kind of attracted much attention recently [18], [96]–[99]. However, non-convex optimization problems into convex optimization because of the involvement of integer variables like RRH problems, which can achieve near optimal solutions. selection indicator and the existence of inter-tier and intra- Specifically, in [101], besides the Branch and Bound tier interference, the joint optimization of multi-dimensional method, the author adopted WMMSE to transform the non- CRRA is always non-convex, making the optimization prob- convex power consumption minimization problem into a con- lem hard to solve. To handle the non-convexity, in general, vex optimization problem with respect to the entries of the there are two kinds of centralized methods. One is using precoding matrix. A WMMSE based precoding scheme to heuristic algorithms which are suitable for finding the op- maximize the downlink system throughput is proposed in [80], timal solutions to NP-hard problems. The other one is to [103]. In [103], the joint power and antenna selection opti- transform these non-convex problems into convex problems mization in a C-RAN is considered. The throughput optimiza- by different methods, and then the primal problems can be tion problem is decomposed into two subproblems. One of the solved by settling the corresponding convex problems. For subproblems, which is non-convex, can be transformed into a the application of heuristic algorithms, in [101], the author convex optimization problem with respect to the entries of the proposed a branch and bound based algorithm to optimize the precoding matrix by WMMSE. In [80], the authors optimize system power consumption with the constraints of transmit the weighted throughput by the optimal user scheduling, RRH delay, transmit power and fronthaul capacity. Although a clustering, and precoding design scheme. Due to the intra- global optimal solution to the problem can be achieved, the tier interference, the objective function is non-convex, and complexity of the algorithm is sufficiently high. Hence, in the it is transformed into a convex form by WMMSE. While in following, we mainly focus on the centralized methods that can [104], the uplink transmission is taken into consideration and a transform non-convex problems into convex problems with a joint downlink and uplink user-RRH association and precoding low complexity. scheme is proposed to minimize the power consumption, in Firstly, Lagrange dual method is effective to design cen- which the joint downlink and uplink optimization problem is tralized algorithms that can optimize resource allocation in transformed into an equivalent downlink problem. As a result, C-RANs when the primal problems satisfy time-sharing con- WMMSE can be used to transform the non-convex downlink dition. In [102], the author investigated the EE maximization problem into a convex problem with respect to the entries of for the heterogeneous cloud radio access network. Due to the precoding matrix. the non-convexity of the primal problem, the author tackled Thirdly, for some precoding design problems in C-RANs, the primal problem by solving its dual problem aiming to l0-norm is incorporated to express RRH selection, which leads minimize the Lagrange function of the primal problem, and corresponding problems to integer programming problems. the dual problem is always convex by definition. Moreover, Moreover, it should be noted that precoding matrices in C- since the primal problem satisfies the time-sharing condition, RANs are sparse because the number of RRHs is much more zero duality gap holds. Thus, the non-convex primal problem than the number of users, which indicates that the percentage IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 17 of the non-zero elements in precoding matrice is low. To 2) Game Model based Optimization: The centralized radio transform such non-convex precoding design problems into resource allocation optimization discussed above requires the convex problems, l1-norm approximation method can be used availability of full and ideal CSIs at the BBU pool, which due to the excellent convex approximation to l0-norm and imposes a significant burden on the fronthaul, especially for the convenience to induce the sparsity of precoding matrices. large-scale C-RANs. To alleviate this burden and decrease In [101], the l1-norm approximation is utilized to transform the computing complexity, one may resort to decentralized the non-convex l0-norm objective function into a convex solutions whereby RRHs self-organize the radio resource form. In [80], the author investigated the reweighted l1-norm allocation and make decisions in a situation of interest conflict approximation to deal with the l0-norm in the constraint and based only on collected local CSIs. To this end, the framework proposed an iterative weight calculation scheme. of game models can be adopted to allocate the radio resources in C-RANs. Detailed investigations of game models in wireless In the l -norm approximation method, each coefficient in 1 networks have also been surveyed in [121], but here we focus the precoding matrix is independent, however, in C-RANs, more on the opportunities of using the game models for CRRA such independence does not always hold. For example, in in C-RANs. As illustrated in Fig. ??, two kinds of game C-RANs, one user is always served by a selected cluster of models, e.g., coalition and contract, have been applied in C- RRHs, which illustrates that the elements not belong to these RANs. RRHs in precoding matrix are set to zero [104]. Besides, For the coalition game model, it is adopted to partition one RRH can be switched off when all of its coefficients in RRHs of C-RANs like BSs in traditional cellular networks into the precoding matrix is set to zero [60]. In these cases, the different clusters called coalitions. In each coalition, RRHs coefficients of precoding matrices should be optimized jointly are cooperative. Specifically, in uplink, authors in [122] made rather than individually, and thus the l -norm approximation 1 multiple single antenna transmitters self-organize into multiple can’t be used directly because the zero entries of the precoding coalitions, where transmitters in the same coalition form a vir- matrices may not align in the same RRH. To overcome tual multi-antenna array. In downlink, RRHs can distributively this challenge, mixed l /l -norm approximation method can 1 p form different coalitions, and interference alignment [123] or be adopted to induce group sparsity. In [60], [104], mixed beamforming [124] can be performed in each coalition to l /l -norm approximation method is proposed to handle a 1 p mitigate inter-cell interference. group sparse based RRH selection problem. The l /l -norm 1 p In C-RANs, RRHs form different clusters to serve UEs, means the l -norm of an l -norm. When the mixed l /l -norm 1 p 1 p and identifying proper schemes to form clusters is essential approximation method is applied in C-RANs, the coefficients to give a high performance of C-RANs. Although the BBU of the precoding matrix will be divided into different groups pool enables centralized optimization, which brings much which can be described as the l -norm. Then execute the p performance gain, a heavy burden can be imposed on the traditional l -norm approximation with respect to these l - 1 p fronthaul links to obtain global CSIs, especially for large- norm groups and the mixed l /l -norm approximation is 1 p scale C-RANs. To reduce transmission burden on the fronthaul completed. In [104], traditional mixed l /l -norm method is 1 p links, it is expected to involve game model based optimization used to transform group sparse based l -norm constraints. In 0 schemes, where RRHs selfishly optimize their own objectives [60], a three stage group sparse precoding design algorithm is only based on local information. In [125], stochastic geometry proposed. The non-convex l -norm constraints are transformed 0 is combined with coalition formation game. First, with a fixed into convex forms by weighted mixed l /l -norm method. 1 p intra-cluster cooperation strategy, authors derived an explicit Further more, the author described the way to design the expression for the successful access probability. Then, a utility weights for different coefficient groups to achieve significant function is defined, which contains the benefit as well as the performance gain. cost incurred by clustering. After that, the problem of RRH In fact, the centralized methods mentioned above are not clustering is modeled as a coalition formation game, and a independent from each other. When solving a non-convex merge-split algorithm is used to cluster RRHs. Cases with and problem, one or more different methods may be utilized. without size limitation of RRH clusters are both considered. For example, in C-RANs, the precoding matrix design and By simulations, it can be seen that the proposed merge-split the RRH selection problem may be optimized at the same based RRH cluster formation algorithm always outperforms time. Thus, the WMMSE method and the norm approximation no clustering strategy since interference is reduced with clus- method can be jointly used to transform the non-convex primal tering. Moreover, the impact of the parameter reflecting the problem into a convex problem. In [80], the non-convex primal cluster formation cost is evaluated. It is found that a lower problem is transformed into the well known quadratically value of the parameter, which means a lower cost for a cluster constrained quadratic programing problem by the WMMSE with fixed size, leads to better performance. This is because method and the norm approximation method. The QCQP lower cluster formation cost leads to larger cluster size, and problem can then be easily solved by some classic methods thus more interference can be eliminated. such as CVX. In [100], [103], the author transformed the non- For the contract game model, it is mainly used to give convex primal problem into a convex problem and then used nodes in wireless networks incentive to reach an agreement the lagrange dual method to transform the convex problem under asymmetric information scenario in H-CRANs. The into a more tractable form. Centralized methods to optimize BBU pool offering the MBS a contract is called a principle, resource allocation in C-RANs are summarized in Table VIII. and the MBS that will potentially accept the contract is IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 18

TABLE VIII SUMMARYOFCENTRALIZEDMETHODSTOOPTIMIZERESOURCEALLOCATIONIN C-RANS

Literature Method Suitable Problems Feature [101] Branch and Bound method Suitable for most kinds of resource allocation problems in C-RANs Optimal solution but with high complexity [102] Lagrange dual method Suitable for problems with zero duality gap Optimal solution [80], [101], WMMSE Suitable for problems whose non-convexity is caused by the inter-tier Sub-optimal solution [103], [104] or intra-tier interference [60], [80], Norm approximation Suitable for problems involving integer variables and being equivalent Sub-optimal solution [101], [104] to a l0-norm called agent. In literatures, contract game model has been individual information and preference of each player [130]. applied to spectrum sharing and relay selection. Specifically, However, the application of matching game model to C- to solve spectrum sharing problem in cognitive radio networks, RANs is still unexplored. One possible application is to use a resource-exchange based contract model and a money- matching game model to solve the RRH association problem. exchange based contract model are adopted in [126] and While many works have applied the matching game model to [127], respectively, with the exact private information about BS association problems [131]–[133]. Each UE can only be secondary UEs unknown to the spectrum owner. In [128], also associated with one BS in these works, which is definitely under an asymmetric information scenario where the source different from the situation in C-RANs. In C-RANs, each is not fully informed about the potential relays, a contract is UE can be associated with multiple RRHs simultaneously, designed to help the source select relays to optimize its own making the RRH association problem hard to be solved. As utility. a result, how to use the matching game model to develop a When the C-RAN is underlaid with MBSs in H-CRANs, low-complexity RRH association algorithm for C-RANs is still mutual interference between the C-RAN and MBSs should be not straightforward and should be researched in the future. A coordinated. Considering the C-RAN and MBSs hold an equal summary of literatures applying game models to C-RANs is position, it is preferable to implement distributed interference shown in Table IX, and a summary of advanced game models coordination. To this end, [129] proposed a contract based is shown in Table X. interference coordination approach for downlink scenario con- taining a C-RAN and an MBS. The BBU pool is taken as a B. Dynamic CRRA with Queue-Awareness principle and the MBS is taken as an agent. The proposed ap- In C-RANs, all RRHs are connected to a BBU pool via proach depends on a cooperative strategy, where three phases fronthaul links and centrally managed by the BBU pool. Due are involved in one time transmission interval with a fixed time to an increased level of frequency sharing and RRH density, length T0. In the first phase with time length t1, RRHs serve all the inter-RRH interferences should be elaborately controlled RRH UEs (RUEs) and the macro UE (MUE), which is named through advanced cooperative signal processing techniques in as RRH-alone with UEs-all scheme. In the second phase with physical (PHY) layer. In addition, the centralized C-RANs time length t2, all RUEs are served by RRHs, and the MUE should be fast adaptive to the changes of traffic demands is not served neither by the MBS nor RRHs. This scheme is caused by users’ joining, leaving, or bursty traffic requests called RRH-alone with RUEs-only scheme. In the last phase, at any time. Therefore, to maximize the performance from RRH-MBS with UEs-separated scheme is performed, where a global perspective, the cross-layer optimization framework RRHs and the MBS serve RUEs and the MUE, respectively. To for cooperative radio resource management (RRM) is essential improve the system performance with the proposed strategy, t1 for future C-RANs. The concept of cross-layer optimization is and t2 should be optimized, aiming at maximizing the utility not new [105], and solving this kind of optimization problems of the BBU pool. Particularly, when the BBU pool knows is known to be computationally complex. However, with the the exact CSIs of the MBS-MUE link, individual rational cloud-computing enabled C-RANs, the time has come to con- constraints should be met, and when the BBU pool only sider the use of advanced cooperative signal processing tech- knows the statistical information about the MBS-MUE link, niques and efficient cooperative RRM algorithms for network both of individual rational constraints and incentive compatible provision. The cross-layer resource optimization in C-RANs constraints should be met. Via simulations, the performance integrates several important cross-layer techniques: 1) ad- gain of the proposed approach compared to FRPC approach vanced cooperative signal processing techniques for PHY layer and TDIC approach is verified. For the FRPC, the same interference management; 2) medium access control (MAC) resource block is reused by the C-RAN and the MBS, and layer algorithms to handle user scheduling, time-frequency the transmit powers of RRHs and the MBS are determined resource allocation, RRH clustering, e.t.c; 3) network layer according to the fairness power control. For the TDIC, the C- solutions to control individual traffic flows and to manage RAN and the MBS transmit separately in different time phases. the network services. Such cross-layer optimization is more Besides the coalition formation game and the contract game challenging than the traditional cellular network optimization model, the matching game model, a Nobel-prize winning as it involves complicated interactions across different layers. framework, is promising as well, since it provides mathe- Lots of articles on cross-layer resource optimization in matically tractable solutions for the combinatorial problem wireless systems have been published recently, most of which of matching players in two distinct sets, depending on the have illustrated that significant SE or energy efficiency EE IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 19

TABLE IX SUMMARY OF LITERATURES APPLYING GAME MODELS TO CLOUD-RANS

Literature Game model Contribution Simulation result [125] Coalition formation Coalition formation game is applied to RRH cluster The performance of clustering strategy is always better game formation, and merge-split algorithm is used to give a than no clustering strategy, and lower cluster formation solution cost leads to better system performance [129] Contract theory Contract theory is applied to interference coordination Significant performance gain is achieved by the proposed between the C-RAN and the MBS, and optimal con- approach compared to FRPC and TDIC schemes tract design is handled under perfect CSIs as well as imperfect CSIs

TABLE X SUMMARY OF ADVANCED GAME MODELS

Game model Feature Advantage Application Coalition formation A set of players form cooperative groups, i.e., Suitable for grouping players of the Inter-cell interference mitigation [123] game coalitions, in order to strengthen their positions in same category with identical utility [124], cluster formation [125] a given situation functions Contract theory The principle offers agents a contract without Suitable for handling cooperation un- Spectrum sharing [126] [127], relay knowing complete agents’ private information, der asymmetric information scenario selection [128], interference coordina- aiming at maximizing its own utility tion [129] Matching theory The players are divided into two sets, and the main Suitable for solving the combinatorial BS association [131] [132] [133] goal of matching is to optimally match players in problem of matching players in two one set with players in the other set, given all play- distinct sets ers’ individual objectives and learned information

3')2G_KX gain can be obtained by the cross-layer optimization of radio BBU Pool resources. The centralized architecture of C-RANs facilitates 79/

Traffic the implementation of CoMP transmission and reception to Traffic Arrival Departure *KRG_G]GXK improve the overall SE. In this context, the cross-layer design )UTZXUR'IZOUT IXUYYRG_KXXGJOU XKYU[XIKGRRUIGZOUT issue of RRH clustering and beamforming to strike the best UVZOSO`GZOUT tradeoff between system throughput and signaling overhead )9/ has been systematically studied in [80]. The cross-layer opti- mization of joint user scheduling and beamforming has been investigated in [106] and [107], where the cross-layer resource ;+ 88. optimization is conducted in a way that maximizes a system 6.?2G_KX utility including throughput gain and proportional fairness. The densification of RRHs in C-RANs gives users the flexibility of Fig. 10. Delay-aware cross-layer radio resource allocation optimization in C-RANs choosing their serving RRHs from a large number of nearby RRHs. This freedom of RRH selection is properly exploited in [103], which achieves substantial gains in terms of both implementation of the algorithms. network throughput and user fairness. A unified way called WMMSE algorithmic framework to design various cross-layer Some developments in this area have been made, and the resource optimization algorithms is proposed in [108], which potential approaches with their advantages and disadvantages can be implemented in a parallel manner. have been identified. The design approaches towards delay- aware cross-layer optimization basically come down to three However, existing literatures [80], [103], [106]–[108] are types: the equivalent rate approach, the Lyapunov optimization typically based on the assumption of infinite queue backlogs approach and the Markov decision process (MDP) approach. and stationary channel conditions. Therefore, only some per- The following surveyed references are by no means exhaustive. formance metrics including the SE, EE, proportional fairness Nonetheless, they still help to present the main roadmap of on- are optimized, and the resulting control policy is only adap- going attempts to design delay-aware cross-layer optimization tive to CSIs. Nevertheless, since the practical C-RANs often algorithms for future C-RANs. operate in the presence of time-varying channel conditions 1) The Equivalent Rate Approach: The principle idea is and random traffic arrivals, and the CSIs-only cross-layer converting the average delay constraints (or objectives) into designs cannot guarantee good delay performance. As shown equivalent average rate constraints (or objectives). The equiv- in Fig. 10, the dynamic cross-layer optimization algorithms alence is usually established based on queuing theory or large should be adaptive to both CSIs and QSI. A delay-aware deviation theory. In [109], the relationship among the average cross-layer optimization framework, taking both the traffic delay requirement, the average arrival rate and the average queue delay and PHY layer performance into account, is service rate is established using the packet flow model of not trivial, as it involves both queuing theory (to model the queuing theory, and the cross-layer optimization problem is traffic queue dynamics) and information theory (to model solved with the equivalent service rate constraints in PHY. In dynamics in PHY). It is made more complicated by practical [110], the statistical delay QoS requirement is characterized issues in C-RANs including fronthaul capacity limitations, by the QoS exponent according to the concept of effective system states (including QSI and CSIs) overhead, and parallel capacity from large deviation theory, and the distributed cross- IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 20 layer resource allocation scheme is proposed to deal with the vestigated in [115] under a unified Lyapunov optimization co-channel interference and guarantee the individual statistical framework, where simulation results demonstrate that the delay QoS requirement. traffic flow control has the advantages of stabilizing traffic Intuitively, the equivalent rate constraint approach provides queues and reducing power consumption. potentially simple solutions for delay-aware cross-layer re- The Lyapunov optimization approach provides a simple source optimization in C-RANs in the sense that the problem alternative to deal with delay-aware cross-layer optimization can be transformed into a pure PHY layer optimization prob- problems. The derived cross-layer control policies are adaptive lem, and the traditional PHY layer optimization approach in to both the CSIs and the QSI. However, for the Lyapunov [106]- [108] can be then readily applied to solve the trans- optimization theory, the queue stability is a weak form of delay formed problem. The optimal control policy is a function of performance, and the average delay bounds are tight when the the CSIs with some weighting shifts by the delay requirements, traffic load is sufficiently high. and hence it is simple to implement in practical communica- 3) The MDP Approach: In C-RANs, the system state is tion systems. However, due to the complex coupling among characterized by the aggregation of CSIs and QSI, and it the queues induced by multi-user interference, it is difficult to evolves as a controlled Markov chain. Thus the delay-aware express delay constraints in terms of all the control actions. cross-layer optimization problem can be modeled as an infinite Therefore, this approach cannot be easily extended to C-RANs horizon average cost MDP, the solution of which can be de- with different levels of interference and fronthaul capacity rived by solving a Bellman equation. However, the cardinality constraints. of the system state space is exponential w.r.t. the number 2) The Lyapunov Optimization Approach: The Lyapunov of queues in C-RANs and hence solving the MDP is quite optimization approach is an extension of Lyapunov drift the- complicated in general. Therefore, the curse of dimensionality, ory, which is originally exploited to analyze the characteristics and the requirements of low complexity and parallel solution of the control policies in the stochastic stability sense [112]. are the two main issues associated with the MDP approach. With Lyapunov optimization, we can stabilize the queues of Furthermore, the existence of multi-user interference makes networks while additionally optimizing performance metrics or the queues of different UEs couple with each other, making satisfying additional constraints across the layers. The delay- the cross-layer optimization more complex. Fortunately, the guaranteed feature of Lyapunov optimization comes from the approximate MDP and stochastic learning methods can help fact that the traffic queues are strongly stable if they have to ensure the low-complexity and parallel implementation of bounded time average backlogs. Using the Lyapunov optimiza- the delay-aware cross-layer resource optimization [116]. tion theory, the procedures of deriving delay-aware cross-layer To achieve the flexible tradeoff between cooperation gain optimization algorithms can be summarized as three main and fronthaul consumption, a hybrid CoMP scheme is de- steps [111]: 1) if needed, transform other average performance signed in [117] for the downlink transmission in C-RANs with constraints into queue stability problems using the technique of capacity-limited fronthaul links and delay-sensitive traffic. To virtual cost queues; 2) choose a Lyapunov function (usually in maintain the low traffic delay, the delay-aware power and quadratic form) and calculate the Lyapunov drift-plus-penalty rate allocation with fronthaul capability and RRH power function; 3) based on the system state observations, minimize constraints are then formulated as an infinite horizon con- the upper bound of Lyapunov drift-plus-penalty function over strained MDP in [117], where the linear approximation, online all polices at each time slot. learning of value functions and stochastic gradient algorithm The Lyapunov optimization has been successfully used to are presented to obtain a low-complexity and parallel solution. derive delay-aware cross-layer optimization algorithms, which However, the online learning of value functions may suffer provides a quantitative method to obtain the desired perfor- from slow convergence and lack of insight. In [118], the delay- mance tradeoff. In [113], a two-stage delay-aware cross-layer aware cooperative beamforming control for delay-sensitive RRM algorithm using the Lyapunov optimization is proposed traffic with limited fronthaul data transfer in C-RANs is in a multi-user multi-input-multi-output (MU-MIMO) systems, considered, which is also formulated as an infinite horizon which can be looked as a general model of C-RANs. The constrained MDP. To tackle the emerging slow convergence proposed cross-layer RRM algorithm consists of a delay-aware of value functions, the continuous time Bellman equation limited CSIs feedback optimization stage over a long timescale is derived and the value function is calculated using the and a user scheduling stage over a short timescale. The well-established calculus theory. However, the unavoidable numerical results in [113] demonstrate a controllable tradeoff interference brought by imperfect CSIT results in some loss between average amount of CSIs feedback and average queue of strict optimality of value function [118]. delay. The Lyapunov optimization theory also has applications The performance degeneration of general MDP-based delay- in the network layer traffic flow control when the exogenous aware resource optimization algorithms caused by the subop- arrival rates are outside of the network stability region. By timal value function is still hard to capture and the tradeoff incorporating the congestion control into the radio resource between optimality and complexity still remains open. It optimization using the lyapunov optimization technique, the should be mentioned that one potential method to calculate the delay-aware cross-layer algorithm is proposed in [114] to optimal value function is based on the perturbation analysis maintain the delay performance and maximize the throughput of value function, as in [119]. However, the calculation is utility for LTE-A networks. The problem of EE-guaranteed extremely complicated and it can only apply to the simple throughput-delay tradeoff in interference-free C-RANs is in- scenario with simple control actions and lightly coupled traffic IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 21 queues. Consequently, an advanced RRH association strategy taking A brief comparison of the three approaches is made in edge cache into account should be exploited. [120], where three different subcarrier and power allocation algorithms in an OFDMA system are proposed. The simulation B. Big Data Mining in C-RANs results demonstrate that the delay performance of the MDP approach is better than those of the equivalent rate constraint In the context of mobile networks, the big data, described by approach and the Lyapunov optimization approach over a volume, variety, velocity, and value, includes subscriber-level, wide operating regime, while the equivalent rate constraint cell-level, core-network-level, and other level data, which can approach and the Lyapunov optimization approach have rela- facilitate the network towards a more proactive one [138]. tively simplified low-complexity solutions. A summary of the Owing to the fast development of big data mining techniques three approaches is listed in Table XI. and the powerful computing capability of the BBU pool, it is feasible to utilize big data technology to extract interesting patterns or knowledge to enhance the self-organizing capa- VII. CHALLENGES AND OPEN ISSUES bilities in C-RANs. For example, by mining the historical In this section, the challenges and benefits brought by the content request data, users’ interests in watching videos can be applications of several attractive techniques in C-RANs are predicted, which helps edge devices cache videos. However, discussed, including edge cache, big data mining, social-aware there are still some technical challenges related to the big data D2D, CR, SDN, and physical layer security. In addition, the mining in C-RANs. For example, the transmission of the large progress of testbed development and trial tests is introduced volume of data collected by edge devices will put a heavy as well. burden on the fronthaul. Furthermore, the computing of sparse, uncertain and incomplete data is a big problem, which requires A. Edge Cache in C-RANs advanced data mining algorithms [139]. Given the exponential increment of the capacity of modern storage units and the consistent decline of the cost per stored C. Social-aware D2D in C-RANs bit [134], it is possible to endow RRHs with large storage D2D communication underlaying cellular system has been capabilities at an acceptable cost. Hence, the burden on the widely researched due to its superiority in improving SE fronthaul links can be alleviated by proactively downloading and EE. To take full advantages of D2D, several challenges the data that will be requested with a high probability and should be handled, including peer discovery, mode selection, then caching it. Moreover, the storage capabilities of UEs can resource allocation, and interference management [140]. Re- be utilized to serve other surrounding UEs as well. In this cently, some works tackle these challenges by leveraging the way, when the requests for the cached data arrive, they will characteristics of social networks, including social tie [141] be served locally, which means the requested data need not be and social community [142]. As a result, the social-aware fetched via fronthaul links. In addition, since the data is more D2D is proposed. For the C-RAN involving social-aware closer to UEs, quality of experience (QoE) can be significantly D2D, much data capacity can be offloaded from fronthaul improved because of a low transmit delay. links, alleviating the fronthaul constraints and shorten the Although the edge cache at edge devices, i.e., RRHs and transmission latency. UEs, brings benefits as mentioned above, there are several Generally speaking, social-aware D2D in a cellular network technical challenges to be handled when edge cache is ap- can be implemented both with and without the assistance of a plied to C-RANs. The first one is the edge cache strategy. BS [143]. The social-aware D2D with the assistance of a BS Specifically, the frequency of data update should be judiciously requires the BS to deliver control signalling, which is hard chosen. High frequency update gives UEs better QoE but to implement in C-RANs since RRHs are mainly deployed to cost more fronthaul resources, and low frequency update is provide high capacity in special zones [29]. Hence, the social- the opposite. Meanwhile, since the popularity of data is the aware D2D without the assistance of a BS is preferred in C- main factor that should be taken into account for the edge RANs. In addition, the selection between D2D mode and C- cache [135], the edge cache strategy under the incomplete RAN mode is critical to achieve high QoS. Furthermore, when information on the popularity is a challenge. The second issue the D2D communication operates in the licensed band, the is data fetching strategy. With edge cache involved in C-RANs, interference between D2D users and RRH users can be severe, the data requested by UEs can now be found in neighbor and meanwhile there can also exist mutual interference among edge devices. Then, if the data requested by a UE is not D2D users. Therefore, how to suppress those interference cached in these devices, an efficient data fetching strategy should be addressed. should be developed to help to decide where to fetch the data and the corresponding route. The last challenge is the RRH association strategy. In [136], [137], some base station D. CR in C-RANs association strategies that consider the impact of edge cache Facing with the ever increasing capacity demands, the spec- have been developed for small cell networks. Nevertheless, trum resource is becoming scarce under the conventional static these strategies are designed for the case where one UE can spectrum allocation policy. In contrast, due to the sporadical only be associated with one base station. While, in C-RANs, usage, a significant amount of assigned spectrum is not fully one UE is often served by multiple RRHs simultaneously. utilized [144]. A promising solution to the spectrum scarcity IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 22

TABLE XI SUMMARY OF APPROACHESTO DYNAMIC CRRA WITH QUEUE AWARENESS

Literature Approaches Solutions Challenges [109], [110] Equivalent rate con- Consider an equivalent problem in PHY by converting the average The solutions are only the functions of CSIs straint approach delay constraints into average rate constraints and cannot be easily extend to C-RANs [111]–[115] Lyapunov optimiza- Consider the resource optimization problem in a stochastic stability The stability is a weak form of traffic delay tion approach sense and obtain an online solution by minimizing the upper bound and the derived solutions may not have good of Lyapunov drift-plus-penalty function performance in the small delay regime [116]–[119] Markov decision Consider the dynamics of the system as a controlled Markov chain, The huge dimensionality of state space and process approach and obtain the solution by solving a Bellman equation together the requirement of low-complexity solution with stochastic leaning or differential equation are the two challenging issues problem is to use the CR technology, which enables secondary C-RAN, the whole C-RAN may collapse in case that this users opportunistically communicate over spectrum holes or controller fails. Secondly, the controller placement should be share the spectrum with licensed users in an under or overlay optimized especially for cases with multiple controllers and manner [145]. the large-scale C-RAN, since the placement has a significant For C-RANs, although centralized LSCP in the BBU pool impact on the processing latency and other network perfor- can enhance SE, CR is preferred to further improve the mance metrics. Thirdly, limited by the service capability of spectrum utilization rate. With the cognitive capability, RRHs the SDN controller, operators need to face with the scalability can interact with radio environment, and find temporally issues in C-RANs. unused spectrum assigned to C-RANs. Together with the CA technique, the operating bandwidth of RRHs can be greatly F. Physical Layer Security in C-RANs enlarged, leading to higher data rates. While, the challenges Due to the broadcasting nature of wireless channels, the mainly lie in the implementation complexity and the cost. For network is vulnerable to various newly arising attacks like example, to achieve spectrum sensing function, the complexity the attack from eavesdroppers. Meanwhile, considering that of transceiver is relatively high [146]. Moreover, to identify encryption often needs considerable computational resources the vacant spectrum, conventionally, the secondary users need and communication overheads, in this context, physical layer to continuously sense the radio spectrum, and hence their security has been an attractive topic recently, which exploits transceivers should always stay in the active state [147], physical characteristics of wireless channels to achieve per- which results in consuming almost the same amount of power fect secrecy against eavesdropping in an information-theoretic as the transmit state. In the future, the structure of using sense [149]. In [150], opportunistic relaying is proposed to CR in C-RANs should be further researched to reduce the reduce the secrecy outage floor of cognitive communications. implementation complexity and improve EE. The security-reliability trade-off of relay-selection is consid- ered in [151], which shows significant improvement compared E. SDN with C-RANs to the classic direct transmission. Recognized as one of the SDN decouples the control plane and data plane, which future wireless networks, the C-RAN should ensure the com- provides a centralized controller and network programmability munication security as well. To deal with this challenge, RRHs [148]. By applying SDN to C-RANs, numerous benefits can can be used to improve the security performance. Specifically, be achieved. For example, the software defined fronthaul in the downlink C-RANs, a UE is typically served by a cluster proposed in [25] can achieve flexible mapping between BBUs of RRHs, and some RRHs can act as jammers to generate and RRHs, which adapts the network to traffic volume and user artificial noise to interfere with the eavesdroppers. However, mobility. Specifically, for a certain region with high mobility some other challenges may arise, and one of them is the users, multiple RRHs are logically mapped to a BBU by SDF selection of jammers among the available RRHs. Furthermore, to realize DAS, reducing handoffs and providing diversity how to design a novel architecture to efficiently support gain. When the region is with static users, SDF will configure physical layer security in C-RANs is still not straightforward. one-to-one mapping to deliver high throughput. In [28], a BBU pool is built in the cloud based on SDN and network function G. Testbed and Trial Test virtualization concepts, in which RRHs are served by the In [13], the notable achievements of C-RANs gained by BBU virtual instances. Similar to the SDN in wired networks, China Mobile in trial tests and testbed development are elab- two controllers are defined: infrastructure manager and service orated. As one of the biggest advocates for C-RANs, China manager. The infrastructure manager is responsible for pooling Mobile has carried out many trial tests in more than 10 cities hardware resources and offering them as service slices to the in China. First, by the TD-LTE based C-RAN field trials, service manager, which consumes service slices according the single fiber bi-direction and CPRI compression solution to the requirement of a BBU virtual instance. By doing so, for the fiber resource saving is shown to have little negative hardware resources in the BBU pool can be utilized more impact on the C-RAN performances. Second, through the trial efficiently, reducing the cost of purchasing new hardware. within China Mobile’s TD-LTE commercial networks in a However, there are still some emerging challenges that dense urban area, it is found that all performance metrics like should be addressed for the application of SDN to C-RANs. throughput and handover successful rate are almost the same Firstly, if only one central controller is in charge of the whole with or without wavelength-division multiplexing which can IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 23 carry dozens of carriers on a single fiber. Third, based on the [8] NTT DoCoMo, “5G radio access: Requirements, concept and technolo- WDM test, the gain of uplink LSCP is demonstrated, reaching gies,” Jul. 2014, Tech. Rep. [9] Nokia Networks, “Liquid radio making radio networks active, adaptive 40-100% in the weak coverage area where reference signal and aware,” 2014, Tech. Rep. received power is lower than -95 dBm. Meanwhile, to validate [10] Ericsson AB, Huawei Technologies, NEC Corporation, Alcatel Lucent, the technical and business feasibility of C-RANs, a GPP based and Nokia Siemens Networks, “Common public radio interface (CPRI); interface specification,” CPRI specification v5.0, Sep. 2011. C-RAN testbed has been developed in China Mobile, which [11] V. Suryaprakash, P. Rost, and G. Fettweis, “Are heterogeneous cloud- can simultaneously support TD-LTE, FDD-LTE and GSM and based radio access networks cost effective?” IEEE J. Sel. Areas deliver a similar performance to the traditional DSP/FPGA Commun., vol. 33, no. 10, pp. 2239-2251, Oct. 2015. [12] O. Simeone, E. Erkip, and S. Shamai, “Full-duplex cloud radio access based systems. To demonstrate the efficiency and achieving networks: An information-theoretic viewpoint,” IEEE Wireless Com- gains of C-RANs in practice, the large-size testbed and the mun. Lett., vol. 3, no. 4, pp. 413-416, Jun. 2014. corresponding trial test for C-RANs are anticipated. [13] C. L. I et al., “Recent progress on C-RAN centralization and cloudifi- cation,” IEEE ACCESS, vol. 2, pp. 1030-1039, Aug. 2014. [14] A. Checko et al., “Cloud RAN for mobile networks-A technology VIII. CONCLUSION overview,” IEEE Commun. Surveys & Tutorials, vol. 17, no. 1, pp. 405-426, Mar. 2015. This paper outlines and surveys the state-of-the-art system [15] J. Wu, Z. Zhang, Y. Hong, and Y. Wen, “Cloud radio access network architecture, key technologies, and open issues in C-RANs. (C-RAN): A primer,” IEEE Netw., vol. 29, no. 1, pp. 35-41, Jan. 2015. With the goal of understanding further intricacies of the key [16] M. Peng, C. Wang, V. Lau, and H. V. Poor, “Fronthaul-constrained cloud radio access networks: Insights and challenges,” IEEE Wireless technologies, we have broadly divided the body of knowledge Commun., vol. 22, no. 2, pp. 152-160, Apr. 2015. into fronthaul compression, large-scale collaborative process- [17] S. Park, O. Simeone, O. Sahin, and S. Shamai Shitz, “Fronthaul com- ing, channel estimation, and cooperative radio resource allo- pression for cloud radio access networks: Signal processing advances inspired by network information theory,” IEEE Signal Process. Mag., cation. Within each of these aspects, we have surveyed the vol. 31, no. 6, pp. 69-79, Oct. 2014. diverse problems and the corresponding solutions that have [18] H. Dahrouj, A. Douik, and O. Dhifallah, “Resource allocation in hetero- been proposed for enabling C-RANs to be configured and geneous cloud radio access networks: Advances and challenges,” IEEE Wireless Commun., vol. 22, no. 3, pp. 66-73, Jul. 2015. operated in a high SE and EE performance manner. [19] T. Pfeiffer, “Next generation mobile fronthaul and midhaul architec- Nevertheless, given the relative infancy of the field, there are tures,” J. Opt. Commun. Network, vol. 7, no. 11, pp. 38-45, Nov. 2015. still quite a number of outstanding problems that need further [20] N. Agata, A. Agata, and K. Nishimura, “A design algorithm for ring topology centralized-radio-access-network,” ONDM, Brest, France, pp. investigation from the perspective of proposing key techniques 173-178, Apr. 2013. and advanced solutions. Given the extensiveness of the re- [21] P. Chanclou, et al., “Optical fiber solution for mobile fronthaul to achieve search areas, it is also concluded that more rigorous investi- cloud radio access network,” Future Network & MobileSummit 2013 Conference Proceedings, Lisboa, pp. 1-11, Jul. 2013. gations are required with greater attention to be focused on [22] K. Ramantas, K. Vlachos, G. Ellinas, A. Hadjiantonis. “A converged transforming the well established fully-centralized C-RAN ar- optical wireless architecture for mobile backhaul networks,” ONDM, chitecture into partially-centralized paradigm with edge cloud Brest, France, pp.155-160, Apr. 2013. [23] “40-gigabit-capable passive optical networks 2 (NG-PON2): computing technology, in which edge cache and social-aware Physical media dependent (PMD) layer specification,” ITU- device-to-device are two key components. Furthermore, with T Recommendation G.989.2, Dec. 2014. [Online]. Available: the introduction of the advanced big data mining, cognitive http://www.itu.int/ITU-T/recommendations/rec.aspx?rec=12097. [24] K. Sundaresan et al., “FluidNet: A flexible cloud-based radio access radio, software defined network, and physical layer security, network for small cells,” IEEE/ACM Trans. Networking, vol. PP, no. 99, the availability of varied degrees of freedom along with the pp. 1-14, Apr. 2015.(DOI: 10.1109/TNET.2015.2419979) associated constraints further beckon the design and validation [25] M. Arslan, K. Sundaresan, and S. Rangarajan, “Software-defined networking in cellular radio access networks: potential and chal- of the original models in the context of C-RANs. The progress lenges,” IEEE Commun. Mag., vol. 53, no. 1, pp. 150-156, Jan. 2015. of trial tests and test bed development of C-RANs can be [26] Y. D. Beyene, R. Jantti, and K. Ruttik, “Cloud-RAN architecture for anticipated to be promoted in the future, which makes C-RANs indoor DAS,” IEEE Access, vol. 2, pp. 1205-1212, Oct. 2014. [27] S. Park, C. Chae, and S. Bahk, “Large-scale antenna operation in commercial rollout as early as possible. heterogeneous cloud radio access networks: A partial centralization approach,” IEEE Wireless Commun., vol. 22, no. 3, pp. 32-40, Jul. REFERENCES 2015. [28] K. Meerja, A. Shami, and A. Refaey, “Hailing cloud empowered radio [1] M. Peng, D. Liang, Y. Wei, J. Li, and H. Chen, “Self-configuration access networks,” IEEE Wireless Commun., vol. 22, no. 1, pp. 122-129, and self-optimization in LTE-Advanced heterogeneous networks,” IEEE Mar. 2015. Commun. Mag., vol. 51, no. 5, pp. 36-45, May 2013. [29] M. Peng, Y. Li, J. Jiang, J. Li, and C. Wang, “Heterogeneous cloud radio [2] M. Peng, Y. Liu, D. Wei, W. Wang, and H. Chen, “Hierarchical access networks: A new perspective for enhancing spectral and energy cooperative relay based heterogeneous networks,” IEEE Wireless Com- efficiencies,” IEEE Wireless Commun., vol. 21, no. 6, pp. 126-135, Dec. munications, vol. 18, no. 3, pp. 48–56, Jun. 2011. 2014. [3] M. Peng, H. Xiang, Y. Cheng, S. Yan, and H. V. Poor, “Inter-tier interfer- [30] M. Peng, Y. Li, Z. Zhao, and C. Wang, “System architecture and key ence suppression in heterogeneous cloud radio access networks,” IEEE technologies for 5G heterogeneous cloud radio access networks,” IEEE Access, Dec. 2015. Netw., vol. 29, no. 2, pp. 6-14, Mar. 2015. [4] W. Huitao and Z. Yong, “C-RAN bearer network solution,” Shenzhen, [31] M. Peng, S. Yan, K. Zhang, and C. Wang, “Fog computing based China, Nov. 2011, Tech. Rep. radio access networks: Issues and challenges.” [Online]. Available: [5] G. Li et al., “Architecture of GPP based, scalable, large-scale C-RAN http://arxiv.org/abs/1506.04233 BBU pool,” IEEE Globecom Workshops (GC Wkshps), Anaheim, USA, [32] J.G. Andrews et al., “What will 5G be? ” IEEE J. Sel. Areas Commun., pp. 267-272, Dec. 2012. vol. 32, no. 6, pp. 1065-1082, Jul. 2014. [6] B. Haberland et al., “Radio base stations in the cloud,” Bell Labs [33] Nokia Networks, “Network architecture for the 5G era,” 2014, Tech. Technical Journal, vol. 18, no. 1, pp. 129-152, 2013. Rep. [7] D. Sabella et al., “RAN as a service: Challenges of designing a flexible [34] M. Maier and B. P. Rimal, “The audacity of fiber-wireless (FiWi) RAN architecture in a cloud-based heterogeneous mobile network,” in networks: Revisited for clouds and cloudlets,” China Commun., vol. Proc. FutureNetwork Summit, pp. 1-8, Jul. 2013. 12, no. 8, pp. 33-45, Aug. 2015. IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 24

[35] Nokia Networks, “Intelligent base stations,” 2014, Tech. Rep. [60] Y. Shi, J. Zhang, and K. B. Letaief, “Group sparse beamforming for [36] A. Dhaini, P. Ho, and G. Shen, “Toward green next-generation passive green Cloud-RAN,” IEEE Trans. Wireless Commun., vol. 13, no. 5, pp. optical networks,” IEEE Commun. Mag., vol. 49, no. 11, pp. 94-101, 2809-2823, May 2014. Nov. 2011. [61] J. Wu, “Green wireless communications: From concept to reality [In- [37] Integrated Device Technology, Inc., “Front-haul compression for emerg- dustry Perspectives],” IEEE Wireless Commun., vol. 19, no. 4, pp. 4-5, ing C-RAN and small cell networks,” Apr. 2013. Aug. 2012. [38] S. Park et al., “Robust and efficient distributed compression for cloud [62] C. R. Berger, Z. Wang, J. Huang, and S. Zhou, “Application of radio access networks,” IEEE Trans. Veh. Tech., vol. 62, no. 2, pp. compressive sensing to sparse channel estimation,” IEEE Commun. 692-703, Feb. 2013. Mag., vol. 48, no. 11, pp. 164-174, Nov. 2010. [39] S. Park, O. Simeone, O. Sahin, and S. Shamai, “Robust layered trans- [63] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. mission and compression for distributed uplink reception in cloud radio 52, no. 4, pp. 1289-1306, Apr. 2006. access networks,” IEEE Trans. Veh. Tech., vol. 63, no. 1, pp. 204-226, [64] E. J. Candes and T. Tao, “Near-optimal signal recovery from random Jan. 2014. projections: Universal encoding strategies? ” IEEE Trans. Inf. Theory, [40] Y. Zhou and W. Yu, “Optimized backhaul compression for uplink cloud vol. 52, no. 12, pp. 5406-5425, Dec. 2006. radio access network,” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. [65] M. Yuan and Y. Lin, “Model selection and estimation in regression with 1295-1307, Jun. 2014. grouped variables,” Journal of the royal statistical society: Series B [41] X. Rao and V. K. N. Lau, “Distributed fronthaul compression and joint (Statistical Methodology), vol. 68, no. 1, pp. 49-67, Feb. 2006. signal recovery in C-RAN,” IEEE Trans. Signal Process., vol. 63, no. [66] S. N. Negahban and M. J. Wainwright, “Simultaneous support re- 4, pp. 1056-1065, Dec. 2014. covery in high dimensions: Benefits and perils of block l1/l∞- [42] S. Park et al., “Joint decompression and decoding for cloud radio access regularization,” IEEE Trans. Inf. Theory, vol. 57, no. 6, pp. 3841-3863, networks,” IEEE Signal Process. Lett., vol. 20, no. 5, pp. 503-506, May. Jun. 2011. 2013. [67] J. Wang, M. Matthaiou, S. Jin, and X. Gao, “Precoder design for [43] J. Kang, O. Simeone, J. Kang, and S. S. Shitz, “Joint signal and channel multiuser MISO systems exploiting statistical and outdated CSIT,” IEEE state information compression for the backhaul of uplink network MIMO Trans. Commun., vol. 61, no. 11, pp. 4551-4564, Nov. 2013. systems,” IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 1555-1567, [68] S. A. Jafar and A. Goldsmith, “Transmitter optimization and optimality Jan. 2014. of beamforming for multiple antenna systems,” IEEE Trans. Wireless [44] H. S. Witsenhausen, “Indirect rate distortion problems,” IEEE Trans. Commun., vol. 3, no. 4, pp. 1165-1175, Jul. 2004. Inf. Theory, vol. 26, no. 5, pp. 518-521, Sep. 1980. [69] X. Gao et al., “Statistical eigenmode transmission over jointly correlated [45] J. Hoydis, M. Kobayashi, and M. Debbah, “Optimal channel training in MIMO channels,” IEEE Trans. Inf. Theory, vol. 55, no. 8, pp. 3735- uplink network MIMO systems,” IEEE Trans. Signal Process., vol. 59, 3750, Aug. 2009. no. 6, pp. 2824-2833, Jun. 2011. [70] J. Wang et al., “Statistical eigenmode-based SDMA for two-user down- [46] L. Liu and R. Zhang, “Optimized uplink transmission in multi-antenna link,” IEEE Trans. Signal Process., vol. 60, no. 10, pp. 5371-5383, Oct. C-RAN with spatial compression and forward,” IEEE Trans. Signal 2012. Process., vol. 63, no. 19, pp. 5083-5095, Oct. 2015. [71] V. Raghavan, S. V. Hanly, and V. V. Veeravalli, “Statistical beamforming [47] S. Jeong, O. Simeone, M. H. Alexander, and J. Kang, “Optimal fronthaul on the grassmann manifold for the two-user broadcast channel,” IEEE quantization for cloud radio positioning,” IEEE Trans. Veh. Tech., vol. Trans. Inf. Theory, vol. 59, no. 10, pp. 6464-6489, Oct. 2013. PP, no. 99, pp. 1-1, May 2015. (DOI: 10.1109/TVT.2015.2431491) [72] M. A. Maddah-Ali and D. Tse, “Completely stale transmitter channel [48] S. Park, O. Simeone, O. Sahin, and S. Shamai, “Multihop backhaul state information is still very useful,” IEEE Trans. Inf. Theory, vol. 58, compression for the uplink of cloud radio access networks,” IEEE no. 7, pp. 4418-4431, Jul. 2012. international symposium on Information theory (ISIT), Honolulu, HI, [73] X. Yi and D. Gesbert, “Precoding methods for the MISO broadcast pp. 2704-2708, Jun. 2014. channel with delayed CSIT,” IEEE Trans. Wireless Commun., vol. 12, [49] S. H. Park, O. Simeone, O. Sahin, and S. Shamai, “Joint precoding no. 5, pp. 1-11, May. 2013. and multivariate backhaul compression for the downlink of cloud radio [74] Y. Shi, J. Zhang, and K. B. Letaief, “CSI overhead reduction with access networks,” IEEE Trans. Signal Process., vol. 61, no. 22, pp. 5646- stochastic beamforming for cloud radio access networks,” in Prco. ICC, 5658, Nov. 2013. Sydney, Australia, pp. 5154-5159, Jun. 2014. [50] S. Park, O. Simeone, O. Sahin, and S. Shamai, “Inter-cluster design [75] Y. Shi, J. Zhang, and K. B. Letaief, “Optimal stochastic coordinated of precoding and fronthaul compression for cloud radio access net- beamforming with compressive CSI acquisition for C-RAN.” IEEE works,” IEEE Wireless Commun. Lett., vol. 3, no. 4, pp. 369-372, Apr. Trans. Signal Process., vol. 63, no. 4, pp. 960-973, Apr. 2015. 2014. [76] N. Li et al., “Robust low-complexity MMSE precoding algorithm for [51] P. Patil and W. Yu, “Hybrid compression and message-sharing strategy cloud radio access networks,” IEEE Commun. Lett., vol. 18, no. 5, pp. for the downlink cloud radio-access network,” Information theory and 773-776, May 2014. applications workshop (ITA), San Diego, CA, pp. 1-6, Feb. 2014. [77] Y. Shi, J. Zhang, and K. B. Letaief, “Robust group sparse beamforming [52] J. Kang, O. Simeone, J. Kang, and S. Shamai, “Fronthaul compres- for multicast green Cloud-RAN with imperfect CSI,” IEEE Trans. sion and precoding design for C-RANs over ergodic fading chan- Signal Process., vol. 63, no. 17, pp. 4647-4659, Sep. 2015. nels,” IEEE Trans. Veh. Tech., vol. PP, no. 99, pp. 1-1, Aug. 2015. [78] B. Dai and W. Yu, “Sparse beamforming for limited-backhaul network (DOI: 10.1109/TVT.2015.2466619) MIMO system via reweighted power minimization,” in Proc. Globecom, [53] N. Jindal and A. Goldsmith, “Dirty-paper coding versus TDMA for Atlanta, America, pp. 1962-1967, Dec. 2013. MIMO broadcast channels,” IEEE Trans. Inf. Theory, vol. 51, no. 5, [79] B. Dai and W. Yu, “Sparse beamforming design for network MIMO pp. 1783-1794, May 2005. system with per-base-station backhaul constraints,” in IEEE 15th [54] M. H. M. Costa, “Writing on dirty paper,” IEEE Trans. Inf. Theory, International Workshop on SPAWC, Toronto, Canada, pp. 294-298, Jun. vol. 29, no. 3, pp. 439-441, May 1983. 2014. [55] K. K. Wang, R. D. Murch, and K. B. Letaief, “A joint-channel diago- [80] B. Dai and W. Yu, “Sparse beamforming and user-centric clustering for nalization for multiuser MIMO antenna systems,” IEEE Trans. Wireless downlink cloud radio access network,” IEEE Access, vol. 2, pp. 1326- Commun., vol. 2, no. 4, pp. 773-786, Jul. 2003. 1339, Oct. 2014. [56] Q. H. Spencer, A. L. Swindlehurst, and M. Haardt, “Zero-forcing [81] S. Boyd et al., “Distributed optimization and statistical learning via the methods for downlink spatial multiplexing in multiuser MIMO chan- alternating direction method of multipliers,” Foundation and Trends in nels,” IEEE Trans. Signal Process., vol. 52, no. 2, pp. 461-471, Feb. Machine Learning, vol. 3, no.1, pp. 1-122. 2011. 2004. [82] T. Kim, F. Sun, and A. J. Paulraj, “Low-complexity MMSE precoding [57] S. Zukang et al., “Sum capacity of multiuser MIMO broadcast channels for coordinated multipoint with per-antenna power constraint,” IEEE with block diagonalization,” IEEE Trans. Wireless Commun., vol. 6, no. Signal Process. Lett., vol. 20, no. 4, pp. 395-398, May 2013. 6, pp. 2040-2045, Jun. 2007. [83] Q. Shi, M. Razaviyayn, Z. Luo, and C. He, “An iteratively weighted [58] H. Kaibin, J. G. Andrews, and R. W. Heath, “Performance of orthogonal MMSE approach to distributed sum-utility maximization for a MIMO beamforming for SDMA with limited feedback,” IEEE Trans. Veh. interfering broadcast channel,” IEEE Trans. Signal Process., vol. 59, Tech., vol. 58, no. 1, pp. 152-164, Jan. 2009. no. 9, pp. 4331-4340, Sep. 2011. [59] Y. Zeng, E. Gunawan, Y. Guan, and J. Liu, “Joint base station selection [84] G. Wang et al., “Acquisition of channel state information in hetero- and linear precoding for cellular networks with multi-cell process- geneous cloud radio access networks: Challenges and research direc- ing,” in Proc. TENCON, Fukuoka, Japan, pp. 1976-1981, Nov. 2010. tions,” IEEE Wireless Commun., vol. 22, no. 3, pp. 100-107, Jun. 2015. IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 25

[85] M. Biguesh and A. Gershman, “Training-based MIMO channel estima- [109] C. Zarakovitis, Q. Ni, D. Skordoulis, and M. Hadjinicolaou, “Power- tion: A study of estimator tradeoffs and optimal training signals,” IEEE efficient cross-layer design for OFDMA systems with heterogeneous Trans. Signal Process., vol. 54, no. 3, pp. 884-893, Mar. 2006. QoS, imperfect CSI, and outage considerations,” IEEE Trans. Veh. Tech., [86] I. Barhumi, G. Leus, and M. Moonen, “Optimal training design for vol. 61, no. 2, pp. 781-798, Feb. 2012. MIMO OFDM systems in mobile wireless channels,” IEEE Trans. [110] J. Tang and X. Zhang, “Cross-layer-model based adaptive resource allo- Signal Process., vol. 51, no. 6, pp. 1615-1624, Jun. 2003. cation for statistical QoS guarantees in mobile wireless networks,” IEEE [87] J. K. Tugnait, “Blind estimation and equalization of MIMO channels Trans. Wireless Commun., vol. 7, no. 6, pp. 2318-2328, Jun. 2008. via multidelay whitening,” IEEE J. Sel. Areas Commun., vol. 19, no. [111] M. J. Neely, “Delay-based network utility maximization,” IEEE/ACM 8, pp. 1507-1519, Aug. 2001. Trans. Networking, vol. 21, no. 1, pp. 41-54, Feb. 2013. [88] Z. Ding and H. V. Poor, “A general framework of precoding design [112] M. J. Neely, Stochastic network optimization with application to for multiple two-way relaying communications,” IEEE Trans. Signal communication and queueing systems, Morgan & Claypool Publishers, Process., vol. 61, no. 6, pp. 1531-1535, Mar. 2013. 2010. [89] J. Tugnait and W. Luo, “On channel estimation using superimposed [113] J. Chen and V. K. N. Lau, “Large deviation delay analysis of queue- training and first-order statistics,” IEEE Commun. Lett., vol. 7, no. 9, aware multi-user MIMO systems with two-timescale mobile-driven pp. 413-416, Sep. 2003. feedback,” IEEE Trans. Signal Process., vol. 61, no. 16, pp. 4067-4076, [90] X. Xie, M. Peng, F. Gao, and W. Wang, “Superimposed training based Aug. 2013. channel estimation for uplink multiple access relay networks,” IEEE [114] H. Ju, B. Liang, J. Li, and X. Yang, “Dynamic joint resource op- Trans. Wireless Commun., vol. 14, no. 8, pp. 4439-4453, Aug. 2015. timization for LTE-Advanced relay networks,” IEEE Trans. Wireless [91] T. Kong and Y. Hua, “Optimal design of source and relay pilots for Commun., vol. 12, no. 11, pp. 5668-5678, Nov. 2013. MIMO relay channel estimation,” IEEE Trans. Signal Process., vol. [115] Y. Li et al., “Throughput-delay tradeoff in interference-free wireless 59, no. 9, pp. 4438-4446, Sep. 2011. networks with guaranteed energy efficiency,” IEEE Trans. Wireless. [92] X. Xie, M. Peng, W. Wang, and H. V. Poor, “Training design and channel Commun., vol. 14, no. 13, pp. 1608-1621, Mar. 2015. estimation in uplink cloud radio access networks,” IEEE Signal Process. [116] W. B. Powell, Approximate dynamic programming: solving the curses Lett., vol. 22, no. 8, pp. 1060-1064, Aug. 2015. of dimensionality. London, U.K.:Wiley-Interscience, 2007. [93] N. Shariati, E. Bengtsson, and M. Debbah, “Low-complexity polynomial [117] J. Li et al., “Resource allocation optimization for delay-sensitive traffic channel estimation in large-scale MIMO with arbitrary statistics,” IEEE in fronthaul constrained cloud radio access networks,” IEEE Syst. J., vol. J. Sel. Topics Signal Process., vol.8, no. 5, pp. 815-830, Oct. 2014. pp, no. 99, pp. 1-12, Nov. 2014.(DOI: 10.1109/JSYST.2014.2364252) [94] S. Zhou, B. Muquet, and G. Giannakis, “Subspace-based blind channel [118] V. K. N. Lau and F. Zhang, “Low complexity delay-constrained beam- estimation for block precoded space-time OFDM,” IEEE Trans. Signal forming for multi-user MIMO systems with imperfect CSIT,” IEEE Process., vol. 50, no. 5, pp. 1215-1228, May 2002. Trans. Signal Process., vol. 61, no. 16, pp. 4090-4099, Aug. 2013. [95] S. Abdallah and I. Psaromiligkos, “Semi-blind channel estimation with [119] F. Zhang and V. K. N. Lau, “Closed-form delay-optimal power control superimposed training for OFDM-based AF two-way relaying,” IEEE for energy harvesting wireless system with finite energy storage,” IEEE Trans. Wireless Commun., vol. 13, no. 5, pp. 2468-2477, May 2014. Trans. Signal Process., vol. 62, no. 21, pp. 5706-5715, Nov. 2014. [96] M. A. Marotta et al., “Resource sharing in heterogeneous cloud radio [120] Y. Cui et al., “A survey on delay-aware resource control for wireless access networks,” IEEE Wireless Commun., vol. 22, no. 3, pp. 74-82, systems - large deviation theory, stochastic Lyapunov drift, and dis- Jun. 2015. tributed stochastic learning,” IEEE Trans. Inf. Theory, vol. 58, no. 3, [97] Y. Shi et al., “Large-scale convex optimization for ultra-dense cloud- pp. 1677-1701, Mar. 2012. RAN,” IEEE Wireless Commun., vol. 22, no. 3, pp. 84-91, Jun. 2015. [121] K. Akkarajitsakul, E. Hossain, D. Niyato, and I.K. Dong, “Game [98] H. Zhang, C. Jiang, and J. Cheng, “Cooperative interference miti- theoretic approaches for multiple access in wireless networks: A sur- gation and handover management for heterogeneous cloud small cell vey,” IEEE Commun. Sur. Tut., vol. 13, no. 3, May-Jun. 2011. networks,” IEEE Trans. Wireless Commun., vol. 22, no. 3, pp. 92-99, [122] W. Saad, Z. Han, M. Debbah, and A. Hjorungnes, “A distributed Jun. 2015. coalition formation framework for fair user cooperation in wireless [99] M. Gerasimenko et al., “Cooperative radio resource management in networks,” IEEE Trans. Wireless Commun., vol. 8, no. 9, Sep. 2009, heterogeneous cloud radio access networks,” IEEE Access, vol. 3, no. pp. 4580-4593. 99, pp. 397-406, Apr. 2015. [123] F. Pantisano et al., “Interference alignment for cooperative femtocell [100] X. Huang, G. Xue, R. Yu, and S. Leng, “Joint scheduling and networks: A game-theoretic approach,” IEEE Trans. Mobile Computing, beamforming coordination in cloud radio access networks with QoS vol. 12, no. 11, Sep. 2012, pp. 2233-2246. guarantees,” IEEE Trans. Veh. Tech., vol. PP, no. 99, pp. 1-1, Aug. [124] S. Guruacharya, D. Niyato, M. Bennis, and D. Kim, “Dynamic coalition 2015 (DOI: 10.1109/TVT.2015.2464278). formation for network MIMO in small cell networks,” IEEE Trans. [101] J. Tang, W. Tay, and T. Quek, “Cross-layer resource allocation with Wireless Commun., vol. 12, no. 10, pp. 5360-5372, Sep. 2013. elastic service scaling in cloud radio access network,” IEEE Trans. [125] Z. Zhao et al., “Cluster formation in cloud radio access networks: Wireless Commun., vol. 14, no. 9, pp. 5068-5081, Sep. 2015. performance analysis and algorithms design,” in Proc. ICC, London, [102] M. Peng et al., “Energy-efficient resource assignment and power UK, pp. 3903-3908, Jun. 2015. allocation in heterogeneous cloud radio access networks,” IEEE Trans. [126] L. Duan, L. Gao, and J. Huang, “Cooperative spectrum sharing: A Veh. Tech., vol. 64, no. 11, pp. 5275-5287, Nov. 2015. contract-based approach,” IEEE Trans. Mobile Computing, vol. 13, no. [103] A. Liu and V. Lau, “Joint power and antenna selection optimization in 1, pp. 174-187, Jan. 2014. large cloud radio access networks,” IEEE Trans. Signal Process., vol. [127] L. Gao, X. Wang, Y. Xu, and Q. Zhang, “Spectrum trading in cognitive 62, no. 5, pp. 1319-1328, Mar. 2014. radio networks: A contract-theoretic modeling approach,” IEEE J. Sel. [104] S. Luo, R. Zhang, and T.J. Lim, “Downlink and uplink energy mini- Areas Commun., vol. 29, no. 4, pp. 843-855, Apr. 2011. mization through user association and beamforming in C-RAN,” IEEE [128] Z. Hasan and V. Bhargava, “Relay selection for OFDM wireless Trans. Wireless Commun., vol. 14, no. 1, pp. 494–508, Jan. 2015. systems under asymmetric information: A contract-theory based ap- [105] B. Fu, Y. Xiao, H. Deng, and H. Zeng, “A survey of cross-layer design proach,” IEEE Trans. Wireless Commun., vol. 12, no. 8, pp. 3824-3837, in wireless networks,” IEEE Commun. Surveys & Tutorials, vol. 16, no. Aug. 2013. 1, pp. 110-126, First Quarter, 2014. [129] M. Peng et al., “Contract-based interference coordination in heteroge- [106] M. Razaviyayn, H. Baligh, A. Callard, and Z. Q. Luo, “Joint user neous cloud radio access networks,” IEEE J. Sel. Areas Commun., vol. grouping and transceiver design in a MIMO interfering broadcast 33, no. 6, pp. 1140-1153, Mar. 2015. channel,” IEEE Trans. Signal Process., vol. 62, no. 1, pp. 85-94, Jan. [130] Y. Gu et al., “Matching theory for future wireless networks: fundamen- 2014. tals and applications,” IEEE Commun. Mag., vol. 53, no. 5, pp. 52-59, [107] M. L. Ku, L. C. Wang, and Y. L. Liu, “Joint antenna beamforming, May. 2015. multiuser scheduling, and power allocation for hierarchical cellular [131] S. Bayat et al., “Distributed user association and femtocell allocation systems,” IEEE J. Sel. Areas Commun., vol. 33, no. 5, pp. 896-909, in heterogeneous wireless networks,” IEEE Trans. Commun., vol. 62, May 2015. no. 8, pp. 3027-3043, Jul. 2014. [108] H. Baligh et al., “Cross-layer provision of future cellular networks: A [132] O. Semiari et al., “Matching theory for priority-based cell association WMMSE-based approach,” IEEE Signal Process. Mag., vol. 31, no. 6, in the downlink of wireless small cell networks,” in Proc. ICASSP, pp. 56-68, Nov. 2014. Florence, pp. 444-448, May 2014. IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 26

[133] W. Saad et al., “A college admissions game for uplink user association Mugen Peng (M’05–SM’11) received the B.E. in wireless small cell networks,” in Proc. INFOCOM, Toronto, pp. 1096- degree in electronics engineering from the Nan- 1104, Apr. 2014. jing University of Posts and Telecommunications, [134] M. Kryder and C. S. Kim, “After hard drives - what comes next?” IEEE Nanjing, China, in 2000, and the Ph.D. degree in Trans. Magnetics, vol. 45, no. 10, pp. 3406-3413, Oct. 2009. communication and information systems from the [135] H. Ahlehagh and S. Dey, “Video-aware scheduling and caching in the Beijing University of Posts and Telecommunications radio access network,” IEEE/ACM Trans. Networking, vol. 22, no. 5, (BUPT), Beijing, China, in 2005. Afterward, he pp. 1444-1462, Jan. 2014. joined BUPT, where he has been a Full Professor [136] F. Pantisano, M. Bennis, W. Saad, and M. Debbah, “Cache-aware with the School of Information and Communica- user association in backhaul-constrained small cell networks,” 12th tion Engineering since 2012. In 2014, he was an International symposium on WiOpt , Hammamet, Tunisia, pp. 37-42, Academic Visiting Fellow with Princeton University, May. 2014. Princeton, NJ, USA. He leads a Research Group focusing on wireless [137] F. Pantisano, M. Bennis, W. Saad, and M. Debbah, “Match to cache: transmission and networking technologies with the Key Laboratory of Uni- Joint user association and backhaul allocation in cache-aware small cell versal Wireless Communications (Ministry of Education), BUPT. His main networks,” in Proc. ICC, London, UK, pp. 3082-3087, Jun. 2015. research areas include wireless communication theory, radio signal processing, [138] A. Imran and A. Zoha, “Challenges in 5G: How to empower SON with and convex optimizations, with a particular interests in cooperative com- big data for enabling 5G,” IEEE Netw., vol. 28, no. 6, pp. 27-33, Nov. munication, self-organization networking, heterogeneous networking, cloud 2014. communication, and internet of things. He has authored/coauthored over 60 [139] X. Wu, X. Zhu, G. Wu, and W. Ding, “Data mining with big refereed IEEE journal papers and over 200 conference proceeding papers. data,” IEEE/ACM Trans. Knowledge and Data Engineering, vol. 26, Dr. Peng was a recipient of the 2014 IEEE ComSoc AP Outstanding no. 1, pp. 97-107, Jun. 2013. Young Researcher Award, and the best paper award in IEEE WCNC 2015, [140] Y. Li et al., “Social-aware D2D communications: Qualitative insights WASA 2015, GameNets 2014, IEEE CIT 2014, ICCTA 2011, IC-BNMT and quantitative analysis,” IEEE Commun. Mag., vol. 52, no. 6, pp. 2010, and IET CCWMC 2009. He received the First Grade Award of the 150-158, Jun. 2014. Technological Invention Award in the Ministry of Education of China for [141] X. Chen, B. Proulx, X. Gong, and J. Zhang, “Exploiting social ties the hierarchical cooperative communication theory and technologies, and the for cooperative D2D communications: A mobile social networking First Grade Award of Technological Invention Award from the China Institute case,” IEEE/ACM Trans. Netw., vol. 23, no. 5, pp. 1471-1484, Oct. of Communications for the contributions to the self-organizing networking 2015. technology in heterogeneous networks. He is on the Editorial/Associate [142] F. Wang, Y. Li, Z. Wang, and Z. Yang, “Social community aware Editorial Board of the IEEE Communications Magazine, IEEE Access, IET resource allocation for D2D communications underlaying cellular net- Communications, International Journal of Antennas and Propagation (IJAP), works,” IEEE Trans. Veh. Tech., vol. PP, no. 99, pp. 1, Jun. 2015. (DOI: and China Communications. He has been the guest leading editor for the 10.1109/TVT.2015.2450774) special issues in the IEEE Wireless Communications, IEEE Access, and IET [143] M. N. Tehrani, M. Uysal, and H. Yanikomeroglu, “Device-to-device Communications. He is the Fellow Of the Institution of Engineering and communication in 5G cellular networks: Challenges, solutions, and Technology (IET). future directions,” IEEE Commun. Mag., vol. 52, no. 5, pp. 86-92, May 2014. [144] I. F. Akyildiz et al., “A survey on spectrum management in cognitive radio networks,” IEEE Commun. Mag., vol. 46, no. 4, pp. 40-48, Apr. 2008. [145] S. Sun, Y. Ju, and Y. Yamao, “Overlay cognitive radio OFDM system for 4G cellular networks,” IEEE Wireless Commun., vol. 20, no. 2, pp. 68-73, Apr. 2013. [146] A. Ghasemi and E. S. Sousa, “Spectrum sensing in cognitive radio Yaohua Sun received his bachelor degree in networks: Requirements, challenges and design trade-offs,” IEEE Com- Telecommunications Engineering with Management mun. Mag., vol. 46, no. 4, pp. 32-39, Apr. 2008. from Beijing University of Posts and Telecommuni- [147] H. Su and X. Zhang, “Energy-efficient spectrum sensing for cognitive cations (BUPT) in 2014 with first class honors, and radio networks,” in Proc. ICC, Cape Town, South Africa, pp. 1-5, May PLACE also got BUPT Excellent Project Award. During the 2010. PHOTO undergraduate study, he won the national scholarship [148] M. Jammal et al., “Software defined networking: State of the art and HERE and the Scholarship of Tang Jun and Sun Chunlan. research challenges,” Computer Networks, vol. 72, pp. 74-98, Oct. 2014. He is currently a Ph.D. student in the Key Laboratory [149] F. Oggier and B. Hassibi, “The secrecy capacity of the MIMO wiretap of Universal Wireless Communications (Ministry of channel,” IEEE Trans. Info. Theory, vol. 57, no. 8, pp. 4961-4972, Oct. Education) at BUPT. His main research interests 2007. include coalition game, Stackelberg game, contract [150] Y. Zou et al., “Securing physical-layer communications for cognitive theory, matching theory, and their applications to resource management in fog radio networks,” IEEE Commun. Mag., vol. 53, no. 9, pp. 48-54, Sep. computing based radio access networks. 2015. [151] Y. Zou, B. Champagne, W. Zhu, and L.Hanzo, “Relay-selection improves the security-reliability trade-off in cognitive radio sys- tems,” IEEE Trans. Commun., vol. 63, no. 1, pp. 215-228, Jan. 2015.

Xuelong Li (M’02-SM’07-F’12) is a full profes- sor with the Center for OPTical IMagery Analysis and Learning (OPTIMAL), State Key Laboratory of Transient Optics and Photonics, Xi’an Institute of PLACE Optics and Precision Mechanics, Chinese Academy PHOTO of Sciences, Xi’an 710119, Shaanxi, P. R. China. HERE IEEE COMMUNICATIONS SURVEYS & TUTORIALS, VOL. X, NO. Y, MON. 2016 27

Zhendong Mao received the B.S. degree in applied physics from the Beijing University of Posts & Communications (BUPT), China, in 2013. He is currently pursuing the Ph.D. degree at BUPT. His PLACE research interests focus on the channel estimation PHOTO and training design in cloud radio access networks HERE (C-RANs) and fog radio access networks.

Chonggang Wang (SM’09) received the B.S. degree from Northwestern Polytechnical University (NPU), Xi’an, China, in 1996, the M.S. degree from the University of Electronic Science and Technology of PLACE China (UESTC), Chengdu, China, in 1999, and the PHOTO Ph.D. degree from the Beijing University of Posts HERE and Telecommunications (BUPT), Beijing, China, in 2002. He is currently a Member Technical Staff with the Innovation Lab, InterDigital Communications, King of Prussia, PA, USA, with a focus on Internet of Things (IoT) R&D activities including technology development and standardization. His current research interest includes IoT, machine-to-machine (M2M) communications, mobile and cloud computing, and big data management and analytics. Dr. Wang was a corecipient of the National Award for Science and Technology Achievement in Telecommu- nications in 2004 on IP Quality of Service (QoS) from the China Institute of Communications. He is the founding Editor-in-Chief of IEEE Internet of Things Journal, an Advisory Board Member of The Institute-IEEE (2015- 2017), and on the editorial board for several journals including IEEE Trans- actions on Big Data and IEEE Access. He has been selected as an IEEE ComSoc Distinguished Lecturer (2015-2016). Virtualized BBU Pool Core network %DFNKDXO

&DSDFLW\ :LUHOHVV :LUHOHVV OLPLWHG:LUHG IURQWKDXO IURQWKDXO KRS   

 



 %%8SRRO 55+  1HWZRUN 8( 0HGLXP$FFHVV &RQWURO 55+  %DVHEDQG 5DGLR)UHTXHQF\ 3URFHVVLQJ :LUHGKRS &35,LQWHUIDFH &35,LQWHUIDFH )URQWKDXO